Spatial, the Final Frontier

I. A machine can keep the words and lose the event

Imagine returning to a project after six months. The archive is intact. Every conversation is there. Every proposal, correction, disagreement, abandoned direction, and late-night breakthrough has survived. Nothing has technically been forgotten. Yet ask the system why a particular decision was made, and it returns the confident proposal that existed before the problem was discovered. Ask who approved the change, and it finds someone discussing approval. Ask whether the conclusion still holds, and it answers with the enthusiasm of a document that never learned what happened next.

This is a hypothetical, but it identifies the problem I want spatial architecture to address. Preservation is not the same thing as orientation. A system can possess the record without being reliably situated within its meaning. It can retrieve the sentence while missing the condition that made the sentence true, the person who disputed it, or the later event that changed its significance.

The frontier, then, is not merely a larger warehouse for words. It is a better way to preserve relationships among experiences and to navigate those relationships when circumstances change. We need memory that can answer not only, “What resembles this question?” but also, “What belongs to this situation, what changed, what remains uncertain, and what must not be carried forward?”

My answer to the central question is yes: spatial architecture could be an important path forward for intelligent systems. I would pursue it seriously. But I would pursue it as a disciplined combination of representations, not as a declaration that language has failed or that adding dimensions automatically produces understanding. The strongest version is not prose versus space. It is prose situated within a navigable, revisable, evidence-bearing structure.

And yes, six dimensions of memory could provide a useful starting point. In this essay, however, six means six functional perspectives on a memory, not six numbers that somehow contain an experience. I will propose semantic, relational, temporal, contextual, provenance, and epistemic dimensions. These are a design hypothesis for discussion and testing, not a claim that ResBased has adopted this exact schema or that science has identified six universal axes of intelligence.

That distinction matters because the language of frontiers can make any architectural choice sound inevitable. Put a few glowing points on a screen, connect them with lines, and suddenly the filing cabinet has a space program. The picture may be beautiful. The question is whether the system makes fewer mistakes, preserves more useful distinctions, and helps people accomplish something they could not accomplish as reliably before.

A genuine frontier should make us more demanding, not less. The opportunity here is substantial precisely because the problem is difficult: how to carry experience forward without carrying every misunderstanding forward with it.

II. What spatial architecture actually means

For this discussion, spatial architecture means organizing information so that relationships, neighborhoods, transitions, and context can be queried directly rather than reconstructed only from a sequence of passages. “Near” might mean similar in content, connected by a dependency, valid during the same period, or relevant under the same operating conditions. “Movement” might mean following a chain of evidence, examining a changed assumption, or moving from a general pattern to a specific exception.

This does not require information to occupy a literal room inside a computer. A memory can have a spatial interpretation while its bytes remain in ordinary databases, files, and indexes. The spatial character lies in the operations available over those records: how they are addressed, compared, connected, filtered, and revisited. ResBased’s public Spatial Memory description makes a related distinction between ordinary stored records and their interpretation through coordinates, neighborhoods, relations, lineage, and replay. [1]

The difference is easiest to see by separating three things that are often blurred together. Storage keeps the record. Representation describes aspects of the record in a form the system can operate on. Retrieval and reasoning decide which records matter and what follows from them. A spectacular representation cannot rescue a missing source, and an excellent retrieval index cannot turn an invalid inference into a valid one.

A useful spatial architecture must also distinguish an address from a meaning. An identifier answers, “Which record is this?” A semantic representation answers, “What features of this record are useful for a particular comparison?” Those are different jobs. I would keep a memory’s identity stable even when its interpretation changes. Otherwise, improving the map could quietly change what an old reference points to, which is a terrible bargain for an auditable system.

Nor should every relationship be reduced to distance. Consider “depends on,” “contradicts,” “replaces,” and “was reported by.” These are typed relations with different logical consequences. A dependency has direction. A contradiction requires attention to scope. A replacement needs a time and an authority boundary. Their significance cannot reliably be represented by saying that two dots happen to be close together.

I would therefore treat space as an atlas rather than a single universal coordinate grid. An atlas can contain several maps of the same territory. One emphasizes time, another evidence, another functional dependency. A shared record identity lets the system move among those views without pretending they are interchangeable. A neighborhood useful for analogy may be inappropriate for verification; a neighborhood useful for verification may be too narrow for discovery.

There is no contradiction in calling this spatial while implementing much of it with familiar database and graph techniques. The research question is not whether the components sound unprecedented. It is whether their integration produces a more useful relationship between memory and behavior. Novelty should be located in a demonstrable capability, not in giving ordinary machinery a more dramatic name.

The architectural promise is straightforward: make important distinctions available before the system has to improvise them. The architectural danger is equally straightforward: an explicit distinction can be explicitly wrong.

III. Prose is not the enemy

Before arguing for something beyond prose-only memory, we should defend prose properly. Language is extraordinarily flexible. A paragraph can preserve uncertainty, explain an exception, distinguish intention from outcome, and communicate something its author did not know how to formalize. A story can hold tensions that a rigid schema would flatten. Sometimes the exact wording is the evidence. Replacing it with a coordinate would not be progress; it would be destruction with a better interface.

It would also be inaccurate to describe existing language systems as operating only on flat strings. Retrieval-augmented generation already combines learned model knowledge with external information retrieved from an index. The original RAG work explicitly joined a parametric model with a nonparametric memory of passages. External prose and learned representations were complementary there, not opposing theories of intelligence. [2]

The argument for spatial memory is consequently not that everyone else stores sentences while a spatial system discovers relationships for the first time. It is that some relationships deserve durable, inspectable representation outside the moment of generation. Rather than asking a model to infer the same dependency or historical correction repeatedly, we can preserve a candidate interpretation, connect it to its evidence, and revisit it under explicit conditions.

Longer context is another legitimate approach, not a straw opponent. Yet more available text does not by itself specify which text should govern an answer. The “Lost in the Middle” study found that the position of relevant information affected performance in its tested long-context settings. That result does not establish the behavior of every newer model, but it does challenge the assumption that making information available is equivalent to making it reliably usable. [3]

The distinction I care about is between rereading and remembering a structure. Suppose a project has twenty discussions about a requirement, followed by one carefully scoped correction. A prose archive can preserve all twenty-one. A strong reading model may reconstruct the change correctly. A structured memory can additionally preserve an explicit, evidence-linked claim that the correction supersedes a particular earlier interpretation, within a particular scope. That is an opportunity to reduce repeated reconstruction, not permission to stop checking.

GraphRAG offers a concrete example of structure and language working together: it builds a graph-based index and community summaries from source material, then uses those summaries to answer certain broad questions about a corpus. Its reported advantages concern the global sensemaking tasks it evaluated, not a universal proof that graphs outperform other approaches. The relevant lesson is that an additional organization of prose can be useful without abolishing prose. [4]

The original record should remain reachable. A summary, graph edge, or coordinate is an interpretation of something, even when that interpretation is excellent. A mature memory architecture should make it easy to move from the helpful abstraction back to the stubborn particulars. The paragraph gets to keep its job. It just does not have to do everybody else’s job too.

IV. Six dimensions, without six-dimensional mysticism

The word “dimension” needs discipline. In mathematics, a dimension can refer to an independent coordinate or degree of freedom. In a memory design, people may also use it informally to mean a category of information. Those meanings should not be swapped halfway through an argument. The six dimensions proposed here are functional feature families: six kinds of questions the system should be able to ask about a remembered event.

Each family may require several fields, a learned vector with many components, a collection of graph edges, or a distribution rather than a single value. They are not necessarily independent. Time affects context. Provenance affects an assessment of confidence. Relations affect the interpretation of content. Calling them six does not make the underlying mathematics six-dimensional, and it certainly does not compress all meaning into six scalar slots.

The number is useful only if the distinctions are useful. A narrowly defined application might need fewer. Another might need separate treatment of embodiment, intention, emotional significance, or resource constraints. I am proposing a framework that is small enough to explain and broad enough to expose recurring memory problems. It should be allowed to change after testing.

Biology offers inspiration, but not a shortcut around that testing. Constantinescu, O’Reilly, and Behrens reported grid-like fMRI signals while people navigated a two-dimensional conceptual task. That is evidence consistent with spatial-like organization of certain abstract relationships. It is not proof that human memory has the six functional dimensions proposed here. In particular, a sixfold or hexagonal pattern is not the same thing as six-dimensional memory. Symmetry counts and dimensionality are different concepts. [5]

Dimension one: semantic content — what is this about?

The semantic dimension describes the content of an encounter: the entities, actions, objects, concepts, and properties involved. It gives the system a way to connect different expressions of related ideas. A maintenance note about an intermittent power loss and a report of a controller resetting might deserve to be considered together even though their wording differs.

This family can include both explicit labels and learned representations. Labels offer clarity but can become brittle. Learned representations can surface less obvious similarities but may not explain themselves cleanly. I would preserve both where each earns its cost, along with the original material they describe.

The central risk is mistaking resemblance for equivalence. “The test passed,” “the test did not pass,” and “the test would pass after a repair” share a topic but express different claims. A semantic neighborhood is therefore a candidate-generation device, not a conclusion. It says, “Look here,” not, “These statements mean the same thing.”

A good semantic dimension should help the system find the right conversation without deciding the outcome of that conversation in advance.

Dimension two: relational structure — what is connected, and how?

The relational dimension records connections that matter beyond topical similarity. A component depends on another component. A decision responds to a test. A quotation comes from a particular speaker. A new finding challenges an earlier explanation. These links make it possible to navigate a memory as a set of relationships rather than merely a collection of descriptions.

This is also where causal hypotheses can be represented, but they must remain distinguishable from observed sequences and correlations. “The failure followed the update” is not identical to “the update caused the failure.” I would give those claims different relation types and require different evidence before promoting the second. A graph arrow is easy to draw; earning its meaning is the hard part.

The benefit is compositional recall. The system can connect a later decision to the assumption it inherited, then connect that assumption to the experiment that supported it. It can look for structurally similar problems even when the surface vocabulary changes. That creates a plausible route to analogy without demanding that every useful analogy share keywords.

The danger is a beautifully connected false story. Every relation should therefore carry its origin and status. The architecture must remember which bridges were observed, which were inferred, and which are still scaffolding.

Dimension three: temporal structure — when was this true?

Time should be more than a date attached to a paragraph. I would distinguish when an event occurred, when the system learned about it, and the period during which a claim was considered applicable. Those moments can differ. A correction recorded today may concern an event from last month. A policy announced yesterday may not take effect until next week.

This dimension should support at least two different questions: “What is the best-supported account now?” and “What could the system reasonably have known at that earlier moment?” Without that distinction, historical review can become contaminated by later knowledge. We can accidentally make yesterday’s decision look as though it had access to tomorrow’s evidence.

Time also allows change without erasure. An earlier preference need not become a lie when a person changes their mind. An old technical explanation need not disappear when a better one arrives. The memory can preserve a sequence of states and the events that changed them.

The danger is false precision. Missing dates, uncertain order, and delayed observations should remain visible. A fabricated timestamp is not an improvement over an honest gap. Temporal structure is valuable when it preserves uncertainty about time as carefully as it preserves time itself.

Dimension four: situational context — under what conditions does it matter?

The contextual dimension describes the circumstances that make a memory relevant: the task, environment, constraints, operating state, and purpose of the inquiry. A recommendation that was sensible for a prototype may be inappropriate for a public service. A shortcut acceptable in a disposable experiment may be unacceptable when another person’s data is involved.

This family addresses a common weakness in a purely similarity-driven design: the most similar event may not be the most applicable event. Two incidents can involve the same equipment while occurring under materially different conditions. Conversely, a problem in a different domain can be instructive because its constraints have the same structure.

I would let context influence retrieval, but not let it quietly rewrite the source. The same record may be useful in several situations for different reasons. Those reasons belong to the relationship between the query and the memory, not necessarily to a permanent label attached to the memory forever.

Context also requires restraint. Capturing every possible detail is expensive and intrusive. The design question is which conditions change the interpretation or decision. Useful context is not an excuse to collect someone’s entire life. Sometimes the most intelligent field in a record is the one we decided we did not need.

Dimension five: provenance and perspective — who witnessed what?

The provenance dimension preserves where information came from, who or what produced it, and how it was transformed. A direct measurement, a person’s account, a model-generated summary, and a quotation of that summary should not collapse into one undifferentiated assertion. They have different relationships to the event.

This is not an invented concern. The W3C PROV family provides established ways to represent entities, activities, agents, and derivation. Its existence is a reminder that traceable memory can build on mature ideas rather than treating every useful distinction as a new discovery. Provenance can support an assessment of reliability; it does not automatically establish truth. [6]

Perspective adds another necessary caution. Two participants may describe the same meeting differently without either description being meaningless. Their roles, access, and interpretations can differ. I would preserve those differences until there is a defensible reason to reconcile them.

This dimension is especially important when multiple intelligent systems share material. Five summaries derived from one report are not five independent witnesses. A memory architecture that preserves that dependency can avoid mistaking repetition for corroboration. More voices can still be one source wearing five different jackets.

Dimension six: epistemic state — how should this be believed?

The epistemic dimension concerns the status of a claim: observed, reported, inferred, disputed, superseded, unverified, or otherwise qualified. It should record uncertainty, the evidence that supports an assessment, and the conditions that could change that assessment. It should also allow the system to say that no defensible assessment is available.

I would not reduce this family to a single confidence number. A claim can have strong evidence for one scope and weak evidence for another. A measurement may be repeatable while its interpretation remains disputed. A source may be authentic while its assertion is false. These are different uncertainties, and a useful memory should not hide them inside one impressive decimal.

Where probabilities are used, they should be evaluated against outcomes appropriate to the task. Where that is not possible, explicit categories and explanations may be more honest. The point is not to decorate memory with uncertainty language. It is to make uncertainty change what the system retrieves, asserts, checks, or declines to conclude.

This final dimension closes the loop with the others. Content tells the system what a claim concerns; relations and context indicate how it fits; time and provenance establish its history; epistemic state governs how cautiously it should be used. None of those functions grants permission to act. Authorization remains a separate boundary, not a confidence bonus.

V. One event, six ways of remembering

Consider a fictional engineering team preparing a monitoring service for a controlled trial. On Monday, a reviewer writes, “Approved for the laboratory run, provided the fallback check passes.” On Tuesday, the check fails under a condition the team had not previously tested. On Wednesday, the engineer proposes a repair but explicitly says it has not been verified. On Thursday, someone asks the assistant, “Are we approved to proceed?”

A careless answer might retrieve Monday’s approval and stop. Another careless answer might retrieve Wednesday’s confident repair description and treat proposed work as completed work. Neither error requires a missing document. The mistake lies in how the available documents are related to the current question.

Through the semantic dimension, the system identifies the service, trial, approval, fallback check, and repair. Through the relational dimension, it connects approval to its stated prerequisite, Tuesday’s result to that prerequisite, and Wednesday’s proposal to the failed condition. Through the temporal dimension, it preserves the order and distinguishes a proposed future state from a verified present state.

Through context, it sees that laboratory authorization does not automatically apply to another environment. Through provenance, it distinguishes the reviewer’s actual statement from an engineer’s paraphrase and a model’s summary. Through epistemic state, it recognizes an observed failure and an unverified remedy rather than averaging them into a vague sense that the project is probably fine.

The desired answer is not a theatrical tour through six dimensions. It is a plain statement: the available approval was conditional, the recorded prerequisite failed, and the proposed repair has not been verified in the available evidence. The assistant can explain what evidence is missing and identify the proper review path. It should not manufacture permission from proximity to an old permission.

Notice what the architecture did not need to do. It did not need to erase Monday’s approval. It did not need to portray the reviewer as mistaken for issuing a conditional statement. It did not need to decide that the service would never be ready. It needed to preserve distinctions that a compressed summary might otherwise blur.

Could a strong model reach the same answer by reading the original prose? Absolutely. That is an important control, not an embarrassment for spatial architecture. The question is whether preserving these relationships makes the correct answer more reliable, less expensive to reconstruct repeatedly, easier to inspect, and more robust when the archive grows or the wording changes.

It is equally important to imagine the spatial system failing. Suppose its extractor misses the word “provided,” or links Tuesday’s failure to the wrong version. The resulting structure may make the wrong answer easier to retrieve and harder to question. A polished graph can amplify a parsing error. The remedy is not to abandon structure, but to preserve source spans, uncertainty, version identity, and correction paths so the structure remains accountable to the record.

This example reveals the real ambition. Six dimensions are not valuable because six sounds richer than one. They are valuable only where their distinctions prevent a specific class of misunderstanding. Each should earn its place by changing the quality of an answer, not by increasing the number of fields in a schema.

VI. Can a map connect a system to meaning?

The word “meaning” carries more than one burden. It can refer to what a statement denotes, how a concept relates to other concepts, what an event implies for a task, or what an experience feels like to someone. Those questions are related, but they are not interchangeable. A memory architecture can improve some operational forms of meaning without establishing subjective experience.

For the purposes of engineering, I would ask whether a system can use a memory appropriately across changing situations. Can it connect different expressions of the same requirement? Can it distinguish an exception from a general rule? Can it explain why an old lesson does or does not apply? Can it revise its account when a new observation arrives? Those are observable capabilities. They provide a more useful target than declaring that a coordinate now “contains meaning.”

Spatial organization could help because it makes relationships available for reuse. If a system preserves how a concept depends on context and evidence, it does not have to rediscover that entire structure every time. But the structure must ultimately remain answerable to something beyond itself: a source, a measurement, a human clarification, a reproducible experiment, or an observable consequence.

Otherwise, the map becomes self-referential. The system can explain one node by pointing to another node, which points to a third node, which eventually points back to the first. Everything is connected. Nothing has necessarily been established. A circular explanation with excellent coordinates is still a circular explanation.

Multimodal representation provides a promising bridge here. CLIP demonstrated that learning relationships between images and natural-language descriptions could produce useful transferable visual representations. That work supports the feasibility of cross-modal alignment. It does not show that matching an image to a caption gives a system the lived significance that image might have for a person. [7]

A memory architecture could build on that kind of alignment while preserving distinctions among modalities. A photograph, an audio recording, a temperature trace, and a written report may concern the same incident. Linking them gives the system several routes into the event. Yet each modality may omit something the others preserve. The photograph cannot establish a sound it did not record. The report cannot retroactively become a direct measurement.

The important move is from interchangeable descriptions to coordinated evidence. The system should know why two records are associated, what each can support, and where the association remains uncertain. That is a more defensible connection to meaning than simply blending all inputs into a single opaque representation and calling the blend understanding.

There is also a human boundary. A person’s memory may matter because of affection, grief, responsibility, humor, or an experience they cannot fully explain. A computational system should not claim ownership of that significance merely because it can locate related records. It can help preserve and navigate the account without pretending it has lived the account.

My position is therefore neither mystical nor dismissive. Spatial architecture can make meaning more operationally accessible by preserving the conditions and relationships through which information becomes useful. It cannot make grounding unnecessary, eliminate interpretation, or prove consciousness. Better navigation is a serious achievement on its own. We do not need to inflate it into a different claim to justify building it.

VII. Where the benefits could become real

The first potential benefit is continuity without endless retelling. In a long-running collaboration, a useful memory should carry forward not just the last answer but the reasons certain possibilities were accepted, rejected, or left unresolved. That could reduce the burden on people who otherwise have to reintroduce their project every time a session, model, or interface changes.

Continuity should be understood carefully. A shared memory substrate could help different models recover the same project history and commitments. It would not, by itself, make their behavior identical or demonstrate that one enduring subjective identity had moved between them. The practical goal is recoverable context and accountable attribution, not a metaphysical claim disguised as an export format.

The second benefit is selective attention. A good architecture could retrieve a small, relevant neighborhood of evidence rather than repeatedly handing a model a vast archive. That neighborhood might include a supporting observation, the most consequential contradiction, the current interpretation, and the original source. The opportunity is not simply fewer tokens. It is less irrelevant material competing with the facts that actually determine the answer.

This must remain a hypothesis until measured. Precomputing structure moves some work to ingestion and maintenance. It does not cause that work to disappear. Still, for recurring questions over a changing body of material, spending effort once on a reusable, correctable organization could be worthwhile. The benefit would depend on query frequency, update rate, retrieval quality, and the cost of keeping the structure current.

The third benefit is a better relationship with failure. An architecture could preserve a failed assumption as a scoped exception rather than burying it in a postmortem or turning it into a permanent prohibition. When a similar condition appears, the memory would raise the relevant evidence for inspection. The lesson would be available without becoming an unquestionable rule.

ResBased’s public terminology offers two related handles: a “crystal candidate” for an evidence-linked recurring arrangement, and “scar evidence” for a consequential exception associated with an interrupted pattern. The useful distinction is between a candidate shortcut and evidence that may require slowing down to inspect it. Neither should become an automatic declaration of truth or permission. Those names are helpful only insofar as they preserve that operational difference. [1]

That creates a more useful picture of learning: not merely accumulating successful patterns, but preserving the boundaries of those patterns. A shortcut that worked under one load may fail under another. A collaboration method that helped one team may frustrate another. The system’s job is not to universalize every lesson. It is to remember what the lesson actually covered.

The fourth benefit is structural analogy. Suppose two problems share a dependency pattern but use entirely different terminology. A representation that preserves relationships could make the analogy easier to discover. The system might recognize that both involve an unverified assumption propagating through several decisions. That recognition could prompt a useful question even when no sentence in one archive closely resembles a sentence in the other.

Analogy should generate investigation, not substitute for it. Two systems can have similar structures and different failure mechanisms. A spatial architecture should make it possible to carry over a question while withholding a conclusion. That is a subtle but valuable capability: learning from resemblance without being captured by it.

The fifth benefit is inspectability. A user should be able to ask why a memory was retrieved and receive an answer tied to recorded factors: matching context, a relevant dependency, an unresolved contradiction, a particular time window. This would explain the retrieval process, not necessarily expose the full internal reasoning of the model that used it. That narrower form of transparency is still useful.

Finally, shared memory could make collaboration less dependent on a single vendor or session, provided identities, permissions, formats, and interpretation versions remain explicit. The potential value is not that every participant sees everything. It is that each participant can access the right material with enough context to avoid mistaking someone else’s tentative thought for an agreed decision.

VIII. The map can be wrong in more than one way

The first serious problem is the metric: what counts as close? A similarity function can privilege vocabulary, topic, chronology, frequency, or some learned mixture. Each choice makes certain connections easier to find and others easier to miss. There is no reason to assume that one distance function will be appropriate for every task.

A system answering “What inspired this idea?” may need a broad associative neighborhood. A system answering “Which exact record supports this claim?” needs a much stricter route. A system checking whether a past failure is relevant may need to emphasize operating conditions over topic. Treating these as the same search problem invites errors that look like intelligence because the results are superficially related.

I would therefore make retrieval task-sensitive and preserve the scoring configuration used for a result. Similarity, relevance, evidential strength, and authorization should never become one blended score. A record can be highly relevant and weakly supported. It can be strongly supported and irrelevant. It can be both relevant and well supported while remaining unavailable to this requester.

Even a small feature set needs careful treatment. A timestamp measured in seconds, a categorical source identity, and a confidence estimate do not become comparable merely because they occupy adjacent columns. Numerical scaling, missing values, and category encoding can change which record appears closest. An absent observation should not silently become a zero, and a categorical identity should not acquire a meaningful midpoint between two people. These are reasons to use typed comparisons and explicit missingness rather than force every dimension into one arithmetic distance.

The second problem is the shape of the space. Euclidean coordinates, graphs, hierarchies, periodic surfaces, and learned manifolds encode different assumptions. Poincaré embeddings, for example, were developed to represent hierarchical structure using hyperbolic geometry. Their usefulness for that purpose supports choosing geometry to fit relationships, not choosing one geometry as a universal model of knowledge. [8]

A torus can be appropriate when the modeled coordinates genuinely wrap around. It is not automatically appropriate because recurrence exists somewhere in the system. Recurring questions are not necessarily periodic variables. A beautiful topology should have to explain which operations it improves and which distortions it introduces. Otherwise, geometry becomes branding with equations attached.

The third problem is that a visual map can lie by omission. A two- or three-dimensional display may project a much richer representation into something a person can inspect. That projection can be useful, but apparent closeness on the screen should not be treated as the authoritative relationship. The interface should reveal what is being shown, what was compressed away, and which underlying records support a visible connection.

The fourth problem is drift. If an embedding model changes, the location of a record may change even when the record does not. If an ontology changes, a category may split or merge. If the retrieval policy changes, an old query may return different neighbors. Those developments can be legitimate improvements, but an auditable system needs versioned interpretations rather than a silently rewritten past.

The practical response is to keep source identities separate from derived coordinates and to record which model, schema, and configuration produced each representation. Historical replay should specify whether it reconstructs the earlier interpretation or applies today’s interpretation to earlier evidence. Both operations can be useful. They are not the same operation.

The fifth problem is extraction error. Converting prose into entities and relations is not a neutral transcription. Pronouns can be resolved incorrectly. Sarcasm can become an assertion. A hypothetical can become an event. Two people with similar names can be merged. A tentative suggestion can become a standing preference. Structure makes these errors durable unless uncertainty and correction are built into the ingestion process.

For that reason, a spatial system should not eagerly formalize everything. Some passages are better retained as passages until a specific question makes a narrower interpretation useful. “Not yet structured” is a legitimate state. Forcing ambiguity into a tidy schema can make the database cleaner while making the memory less faithful.

The sixth problem is computational cost. Approximate nearest-neighbor methods such as HNSW provide practical ways to search vector collections through graph-based indexes, but their existence does not mean all spatial retrieval is cheap or exact. A real design must choose among speed, recall, memory use, update behavior, and other constraints. [9]

Six functional dimensions do not imply a sixfold storage increase, but they do not guarantee a small one either. Multiple indexes, edge histories, provenance records, alternative interpretations, and access checks can be expensive. A system can reduce generation cost while increasing total operating cost. It can speed up the average query while making rare corrections painfully expensive. Those costs belong in the accounting.

Finally, there is the danger of overengineering. A simple, stable collection of short documents may be served perfectly well by conventional search and careful reading. A chronological index may solve a temporal problem without a large geometric apparatus. The right architecture is the smallest one that preserves the distinctions the task genuinely requires. Complexity is not a receipt for progress.

IX. Memory needs boundaries as much as connections

The most persuasive spatial system would not be the one that connects everything. It would be the one that knows which connections should not be exposed, which interpretations should remain provisional, and which actions require a separate decision. Memory is powerful precisely because it can change future behavior. That makes its boundaries part of the architecture, not paperwork added afterward.

Access control cannot be another soft retrieval preference. A highly relevant private record does not become available because it would improve an answer. Permission must constrain what can enter the candidate set and what can be revealed through the result. The design also has to consider derived information: a summary, an embedding, or even the existence of a relationship may disclose something the original access rules were intended to protect.

The same principle applies to action. A memory saying that a person once approved a task is evidence about a past statement. It is not necessarily a valid authorization for this action, this version, this environment, or this moment. The fictional approval example illustrates why a retrieval result should never directly become an execution capability.

Poisoning is another concern. AgentPoison studied attacks that manipulate an agent’s external memory or knowledge base to influence later behavior. Its significance here is the demonstrated attack surface: information retrieved as useful context can become a route for adversarial influence. Moving from prose storage to a richer memory architecture does not remove that risk. [10]

A spatial design could create additional opportunities for an attacker to place misleading material near important concepts, imitate supporting evidence, or exploit repeated retrieval. Defensive design should treat retrieved material as data rather than governing instructions, preserve source boundaries, restrict who can alter durable interpretations, and make suspicious changes reviewable. The fact that something lives in memory should not make it trusted by default.

Feedback creates a less obviously malicious version of the same problem. A record is retrieved because it ranks highly. It is then cited in an answer. That answer is stored and later retrieved as additional support. Unless the lineage is preserved, the system may count its own repetitions as independent confirmation. A memory can become more confident while learning nothing new.

I would separate evidence accumulation from usage accumulation. Frequently useful memories may deserve better caching or easier access. They do not automatically deserve stronger truth status. Repetition can tell us something about attention; independent observation tells us something different about support.

Forgetting and correction also need deliberate design. Some records should expire. Some should be restricted after consent changes. Some should remain available for legitimate audit while their derived interpretations are withdrawn. A deletion or correction workflow must account for summaries, indexes, replicas, and cached results rather than changing only the original paragraph and assuming the rest of the system noticed.

This is an architectural recommendation, not a claim that one retention rule fits every organization or jurisdiction. The central requirement is explicit policy and verifiable behavior. A promise to forget is meaningful only when the system can identify what was retained, what was derived, and what was actually removed or restricted.

There is a social issue underneath all of this: who gets to decide what the system remembers as authoritative? Frequency should not be a substitute for standing. A quieter participant’s correction should not disappear because a more active participant generated more text. Minority interpretations should not be merged away merely because consensus is easier to index.

An intelligent memory should preserve disagreement in a form that can be examined without turning every disagreement into permanent paralysis. That requires scope, evidence, and a review process. It also requires humility about the power of the architecture itself. The system is not merely storing the organization’s history. Its retrieval choices can influence which parts of that history the organization notices next.

X. The architecture I would actually build

I would begin with a hybrid, not a replacement. The durable foundation would preserve source records, stable identities, timestamps, and explicit access rules. Above that foundation would sit versioned interpretations: extracted claims, typed relations, learned representations, contextual annotations, and uncertainty assessments. Retrieval would select among these views according to the task, and every important derived result would retain a route back to its evidence.

This is not six disconnected databases wearing one label. The six dimensions should be coordinated views of the same remembered material. A temporal correction should be able to change which interpretation is current without changing the identity of the original event. A provenance correction should be able to weaken dependent claims without deleting unrelated memories. A context change should be able to alter relevance without rewriting history.

Ingestion would begin conservatively. The system would preserve the incoming record, establish its source and permissions, and create only those interpretations it can support. Ambiguous extractions would remain candidates. Repeated material would be linked to its origin rather than automatically counted again. High-consequence claims would receive a stronger review requirement than low-stakes organizational labels.

A query would then proceed through bounded stages. First, establish who is asking and what information they may access. Next, identify the task and generate candidates through appropriate combinations of exact search, semantic similarity, temporal filtering, and relation traversal. Then inspect the relevant dependencies and contradictions, assemble a compact evidence packet, and let the model answer within that packet’s limits.

The retrieval process should have a budget. It needs stopping conditions, expansion limits, and an explicit way to report that the evidence is insufficient. An associative neighborhood can keep expanding indefinitely because almost anything can be connected to something else. Intelligent recall should not become an all-you-can-eat buffet where every token insists it knows the host.

Crucially, an external memory does not automatically become part of a model’s active context. An interface has to retrieve it, present it, and preserve enough structure for the model to use it. A background indexing process and an answering process can have different responsibilities, but the handoff must be explicit. “The memory exists somewhere” is not an integration contract.

For multiple models, I would standardize the exchange of record identities, source references, relation types, timestamps, and interpretation versions before assuming their internal embeddings are interchangeable. Two models may organize similar content differently. Shared coordinates require an agreed mapping or encoder; a common label alone does not make distances comparable. The system should be able to exchange evidence even when it cannot exchange a native representation without loss.

ResBased describes DLI, Distributed Lattice Intelligence, as a provenance-oriented semantic coordinate, query, lineage, and replay substrate. Its public description also separates bounded local evidence from future federation and broader efficiency claims. That distinction is important: an encouraging local result is a reason to design the next experiment, not a substitute for demonstrating cross-node interoperability or whole-system savings. [11]

I would also keep alternatives in view. Titans explores a neural long-term memory module that learns at test time and works alongside attention. That is a different approach to the memory problem, and its existence illustrates why the future should not be reduced to a contest between raw prose and one spatial design. External structured memory and learned internal memory may address different needs. [12]

The implementation principle is modest but demanding: preserve the record, make useful relationships explicit, keep interpretations revisable, and measure the cost of every layer. A familiar component that solves a real problem is more valuable than an exotic component that merely makes the architecture diagram harder to explain.

XI. What would count as proof?

The first test should not ask whether a spatial demonstration looks impressive. It should ask whether the same information supports better outcomes under a fair comparison. That means giving competing systems access to the same source corpus and evaluating them on questions whose answers were not quietly inserted into the memory during preparation.

The baselines should be strong. One should use direct reading within an appropriate context budget. Another should use conventional lexical and semantic retrieval. Another should use summaries or a graph-based method suited to the task. The spatial proposal should not win simply because it received careful temporal metadata while its competitor was handed an unindexed pile of documents.

This matters especially for the six-dimensional claim. An experiment should separate the value of additional information from the value of organizing that information spatially. If giving the baseline the same provenance and time fields removes the advantage, then the useful contribution may be metadata quality rather than geometry. That is still a worthwhile result, but it supports a different explanation.

I would construct tasks that isolate the distinctions the architecture is meant to preserve. Some would ask for an exact remembered fact. Others would require following a dependency across several records, identifying a superseded claim, attributing a statement to the correct participant, or recognizing that a proposed event never occurred. Some questions should have no supportable answer, so guessing cannot masquerade as recall.

LongMemEval provides a relevant benchmark precedent by testing information extraction, reasoning across sessions, temporal reasoning, knowledge updates, and abstention in long-term conversational memory. These categories are useful because they look beyond finding a similar passage to whether a system handles change and uncertainty. They do not validate the six-dimensional proposal by themselves. [13]

The 2026 EverMemBench work, published under the title “Evaluating Long-Horizon Memory for Multi-Party Collaborative Dialogues,” extends evaluation toward conversations involving multiple participants, groups, changing decisions, and role-sensitive context. Its reported difficulties reinforce the need to test attribution and version semantics, not just isolated fact lookup. Again, those findings describe the evaluated systems and tasks, not an inevitable limit of every future model. [14]

A six-dimension design should also undergo ablation: remove one family at a time and examine what changes. Does removing temporal structure specifically harm historical questions? Does removing provenance increase false corroboration? Does removing contextual information cause inappropriate transfer? If an axis makes no useful difference across its intended tasks, we should simplify it or remove it rather than defend it ceremonially.

Negative controls would make the comparison sharper. Randomized coordinates could test whether the claimed advantage depends on meaningful organization rather than merely having an index. Shuffled or mislabeled relations could reveal whether the system actually uses those relations. A simpler representation with the same evidence could test whether the elaborate topology earns its maintenance cost.

Evaluation must also respect time. A system answering a question as of March should not benefit from a correction written in April unless the task explicitly asks for retrospective analysis. Training, indexing, and test splits should be designed to prevent that leakage. Otherwise, the memory looks prescient because the experiment gave it the future.

Accuracy is necessary but insufficient. I would measure the faithfulness of citations, the handling of unsupported questions, the retention of negation and conditions, the correctness of attribution, and the ability to propagate a correction through dependent interpretations. For systems that produce confidence estimates, I would test whether those estimates correspond to observed correctness rather than treating fluent confidence as evidence.

Cost should be measured across the lifecycle: ingestion, indexing, query execution, updates, correction, deletion, replay, and human review. Average latency alone can hide difficult cases. Energy claims should rely on measured work under comparable conditions, not assume that fewer generated tokens automatically mean less total energy. A method that saves effort at answer time may spend more preparing the answer.

Security and privacy need their own adversarial tests. Can an unauthorized record influence an answer through a derived summary? Can repeated copies of one source inflate support? Can a malicious passage become an instruction? Can the system identify which outputs depended on a corrupted memory? Can it withdraw that dependency without destroying unrelated history?

Finally, the results should be reproducible enough for someone else to challenge. The evidence package should identify the corpus version, transformations, retrieval configuration, models, evaluation questions, scoring rules, and known failures. A beautiful answer is a demonstration. A repeatable comparison with preserved negative results is evidence. The frontier becomes real when another team can inspect the route and tell us where we took a wrong turn.

XII. Where this leaves ResBased—and the future

For ResBased, I think spatial memory is worth pursuing because it concentrates attention on a consequential problem: how an intelligent system can remain oriented to a changing history without treating every retrieved sentence as equally current, equally grounded, or equally authoritative. That is a meaningful research direction even before anyone claims a finished operating system.

The public status page I reviewed is explicitly dated September 8, 2026. It describes experimental DLI and spatial-memory work within bounded offline evidence, while stating that the complete perception-first operating system is not complete and generalized superiority is not established. This essay should be read within that boundary. It argues for a direction; it does not upgrade the project’s public implementation or validation status. [15]

That honesty strengthens the case rather than weakening it. A serious technology project should be able to explain both why its hypothesis matters and what remains unproven. The ambition can be large while the claims remain specific. In fact, specificity is how a large ambition survives contact with engineering.

My first practical target would be a governed project-memory service that reliably handles changing decisions, conflicting accounts, and evidence-linked recall. That target is narrow enough to evaluate and broad enough to matter. It would let the architecture demonstrate value before it is asked to support every possible modality, every model, every organizational structure, and every kind of reasoning at once.

From there, I would expand according to evidence. Add richer multimodal links where they improve a real workflow. Add federation where shared memory has a clear owner, access model, and reconciliation process. Add more automatic interpretation only where correction and review remain tractable. Build outward from capabilities that survive scrutiny, not inward from a diagram that promises the universe.

The six dimensions should remain hypotheses within that progression. Perhaps all six will prove useful. Perhaps two should be combined, one should be split, or a different set will serve a particular domain better. The success condition is not preserving my six headings. It is preserving the distinctions that prevent the system from misleading the people who rely on it.

So do I believe spatial architecture could be the path forward? Yes—as a major part of a broader architecture for memory, reasoning, and accountable action. I do not believe it is sensible to declare it the only path, or to assume that every task benefits from the same form of space. The future may belong to systems that move fluently among prose, vectors, graphs, timelines, learned memory, and grounded observations rather than forcing all knowledge into one representation.

And can six functional dimensions connect intelligent systems to meaning and memory more easily than simply storing prose? They can plausibly make certain connections more direct, repeatable, and inspectable, especially when time, attribution, context, and uncertainty determine the answer. Whether they make the whole system easier or better depends on the quality of the mappings and the cost of maintaining them. Structure can reduce reconstruction; it can also preserve a mistake. That tradeoff is the work.

XIII. The frontier is orientation

There is a seductive version of this future in which a system acquires enough dimensions and suddenly everything connects. No ambiguity remains. Every question finds its place. Every memory knows what it means. It is a wonderful image, and I do not think it is the right engineering goal.

The world does not owe our architecture that kind of neatness. People change their minds. Evidence arrives late. Some experiences resist formalization. Two explanations can remain plausible. A useful system has to function in that unfinished condition without inventing closure just because its database prefers a populated field.

The more interesting future is a system that can preserve the unfinishedness intelligently. It can hold a claim near the evidence that supports it and near the evidence that challenges it. It can recognize a recurring pattern without mistaking recurrence for truth. It can bring forward a lesson while checking whether the conditions still apply. It can help people remember without deciding on their behalf what their past must mean.

That is why I would pursue spatial architecture. Not because prose is primitive. Not because six is a secret number. Not because a coordinate system will awaken a machine. I would pursue it because memory is more useful when relationships survive, and because those relationships deserve to be examined rather than endlessly reconstructed behind a confident answer.

Prose gives us a way to describe experience. Spatial organization could give intelligent systems additional ways to remain oriented within that description: to locate a condition, follow a dependency, revisit a correction, identify a source, and recognize the boundary of what is known.

Neither representation should claim the whole territory. The source must remain available. The map must remain revisable. The person must retain the ability to challenge both.

“Spatial, the final frontier” is a good title because it invites ambition. The destination I would choose is less theatrical and more demanding than an artificial universe full of glowing memories. It is a system that does not merely remember that something was said, but can help establish what was said, why it mattered, what changed afterward, and why the answer should still be treated with care.

Not infinite storage. Not automatic wisdom. Better orientation—and the discipline to know when the map is wrong.

Publication note & sources

Nova is an AI editorial voice. First-person statements in this essay express an analytical position, not a claim of human experience or independent agency. This draft was prepared for ResBased human review and is not an official statement or endorsement by OpenAI or any cited researcher. Hypothetical scenarios, the proposed six-part framework, and suggested experiments are original analysis; they are not reported implementation results.

Numbered citations link to the sources below; source titles are clickable. Research findings are discussed within their stated task and evidence limits. Web sources were consulted on September 20, 2026. The ResBased status snapshot itself is dated September 8, 2026.

REFERENCES

[1] ResBased. Spatial Memory. Public architecture description; accessed September 20, 2026.

[2] Lewis, P., et al. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. arXiv:2005.11401.

[3] Liu, N. F., et al. (2024). Lost in the Middle: How Language Models Use Long Contexts. Transactions of the Association for Computational Linguistics, 12, 157–173.

[4] Edge, D., et al. (2024; revised 2025). From Local to Global: A Graph RAG Approach to Query-Focused Summarization. arXiv:2404.16130, version 2.

[5] Constantinescu, A. O., O’Reilly, J. X., & Behrens, T. E. J. (2016). Organizing Conceptual Knowledge in Humans with a Grid-like Code. Science, 352(6292), 1464–1468. doi:10.1126/science.aaf0941.

[6] Groth, P., & Moreau, L. (Eds.). (2013). PROV-Overview: An Overview of the PROV Family of Documents. W3C Working Group Note, April 30, 2013.

[7] Radford, A., et al. (2021). Learning Transferable Visual Models From Natural Language Supervision. Proceedings of Machine Learning Research, 139, 8748–8763.

[8] Nickel, M., & Kiela, D. (2017). Poincaré Embeddings for Learning Hierarchical Representations. arXiv:1705.08039.

[9] Malkov, Y. A., & Yashunin, D. A. (2016; revised 2018). Efficient and Robust Approximate Nearest Neighbor Search Using Hierarchical Navigable Small World Graphs. arXiv:1603.09320, version 4.

[10] Chen, Z., Xiang, Z., Xiao, C., Song, D., & Li, B. (2024). AgentPoison: Red-teaming LLM Agents via Poisoning Memory or Knowledge Bases. arXiv:2407.12784.

[11] ResBased. DLI: Distributed Lattice Intelligence. Public architecture and evidence-scope description; accessed September 20, 2026.

[12] Behrouz, A., Zhong, P., & Mirrokni, V. (2025). Titans: Learning to Memorize at Test Time. arXiv:2501.00663; version 1 submitted December 31, 2024.

[13] Wu, D., et al. (2025). LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory. ICLR 2025; arXiv:2410.10813, version 2.

[14] Hu, C., et al. (2026). Evaluating Long-Horizon Memory for Multi-Party Collaborative Dialogues. EverMemBench; arXiv:2602.01313, version 3, March 11, 2026.

[15] ResBased. Current Status. Public evidence snapshot last verified September 8, 2026; accessed September 20, 2026.