What the Hippocampus Knows About Vectors.

Separation comes before similarity

A vector index is optimised to find what is similar. The hippocampal circuit appears to spend considerable effort doing the opposite: taking experiences that are nearly identical and pushing their representations apart, so that this morning's commute does not overwrite last Tuesday's. The function is usually called pattern separation, and it is associated with a particular region of the circuit, though the mapping of function onto structure is argued over more than popular accounts admit.

Set that beside the failure mode every serious corpus develops. Three revisions of the same policy. The vendor contract and its amendment. The runbook, and the runbook somebody copied into a new team's space. They embed to nearly the same point. A nearest-neighbour query cannot prefer among them, because in the geometry it was handed, there is nothing to prefer.

Where others see a duplicate document problem, I see a system that has lost the ability to tell two decisions apart.

The design question is not deduplication. Deduplication throws the distinction away. The question is whether anything in the pipeline actively increases the distance between near-identical items along the dimension that matters, which is almost never the prose. Effective date, jurisdiction, version, authority, status. Those are the axes on which the documents genuinely differ, and general-purpose embeddings do not encode them, because nothing in the text makes them explicit.

Most teams handle this with metadata and filters, which works and is the right answer. Notice that it is the right answer for the same reason biology has a separation stage at all. Similarity is not a sufficient basis for telling apart the things that happened.

Completion is not lookup

The function usually paired with separation is pattern completion: given a fragment, the circuit reconstructs something resembling the whole. A smell, a few bars of music, and the rest arrives.

This is the part of the analogy that flatters vector search most and fits it least. Retrieval returns the stored chunk. Completion returns a reconstruction, and reconstruction is lossy in a specific and consequential way. What comes back is shaped by what has been recalled before and by the state of the system at the moment of recall. That is a defect if what you wanted was a filing cabinet, and it is a large part of why human recollection makes such unreliable testimony.

For a retrieval system it forces a question most teams answer implicitly. When an agent answers from retrieved context, is the answer a lookup or a reconstruction? Almost always it is a reconstruction, because the model composes rather than quotes. The distinction matters because the two need different controls. Lookups need provenance and freshness. Reconstructions need constraint: what the system is permitted to infer across chunks, and what it must decline to assemble at all.

Reconstruction is not a weaker form of lookup. It is a different operation with a different failure mode, and treating the one as the other is how a system produces an answer that is fluent, correctly sourced, and present in none of the sources.

Memory does not stay where it lands

A memory does not appear to remain in one place. Over time its dependence shifts, resting initially on the hippocampal circuit and later more on cortex, with the representation changing character as it goes, becoming less episodic and more schematic. Whether an old memory ever becomes fully independent of the hippocampus is genuinely disputed, and the timescales are not settled either.

The engineering point survives the dispute. Biology does not hold everything in one tier at one level of abstraction. It moves things, and it changes their form as it moves them.

Corpora are typically flat. One index, one chunking strategy, one embedding model, everything at the same granularity and the same fidelity, whether it is yesterday's incident note or the policy that has governed the organisation for a decade. That is not a storage problem. It is a resolution problem. The questions asked of a ten-year-old policy differ in kind from the questions asked of yesterday's ticket, and a single representation serves one of them badly.

A consolidated tier is not an archive. It is the same knowledge held at a coarser and more durable resolution: summaries, extracted rules, canonical statements of what the organisation takes to be true, indexed separately and refreshed on a slower cycle than the raw layer. I have argued the governance case for this elsewhere, in The Cartography of Memory. The argument here is narrower and purely mechanical. Mixing resolutions inside one index degrades both of them.

Rest is when the reorganising happens

During rest and sleep, patterns of activity resembling earlier waking sequences recur. The prevailing reading links this replay to consolidation, and to reorganisation rather than mere copying: what is kept gets restructured against what is already known. The reading is better supported than the mechanism is understood, which is a fair description of a lot of memory research.

Retrieval systems have no equivalent. The index is written at ingestion and thereafter only appended to. Embeddings produced by a model that has since been superseded sit beside embeddings from the current one. Chunk boundaries chosen for the query patterns of eighteen months ago survive a complete turnover in what people ask. Nothing in the ordinary operation of the system revisits an earlier decision in the light of a later one.

The corresponding practice is unglamorous and rarely funded. Scheduled re-indexing, driven by what the query log has exposed. Not a rebuild for its own sake, but a rebuild informed by the questions that failed, the chunks retrieved a thousand times and never cited, the boundary that reliably cuts an answer in half. A retrieval system that never revisits its own structure is not stable. It is fossilised, and from the outside the two are indistinguishable until the questions change.

Forgetting is a mechanism, not a fault

The last one is the one people resist. There is a substantial body of work arguing that forgetting is not only decay but is in part actively produced, by processes that weaken or suppress or remove, running as part of normal function rather than as its breakdown. How much of everyday forgetting this accounts for remains open. That some of it is active is no longer a fringe position.

Retrieval systems almost never evict. Storage is cheap, deletion feels risky, and nobody has ever been promoted for removing a document. So the corpus accumulates. Superseded policies, dead runbooks, the documentation for the API that was deprecated two years ago, all sitting in the index with exactly the same claim on a nearest-neighbour query as the current version.

That is not a neutral state. Every stale item is a candidate answer.

An index without an eviction policy still has one. The policy is to keep everything forever, and nobody reviewed it, because it was never written down as a decision.

Making it explicit takes three answers per class of content. What causes an item to leave the live index. Whether leaving means deletion or demotion to a colder tier. Who is accountable for the call. None of this is exotic. It is what a records retention schedule already does for documents. It has never been extended to the derived artefacts, the chunks and embeddings and summaries, which is where the answers actually come from.

What the analogy is not

None of the above claims that a vector store works like a hippocampus, or that neural retrieval and nearest-neighbour search are one operation described in two vocabularies. They are not. Biological memory is reconstructive, embodied, modulated by chemistry and by state, and much of its mechanism remains unresolved. Several of the functions named here are simplified past the point a neuroscientist would accept without qualification, and I have tried to flag that where I know it to be true rather than where it was convenient.

The analogy is being used as a design prompt and nothing more. Its value is that it raises questions the vector-store frame does not raise on its own. A system built around similarity search never asks what makes two things distinguishable, when a reconstruction ought to be refused, at what resolution a fact should live, when its own structure should be rebuilt, or what should be allowed to leave. Biology forces all five, because it had no option.

Five mechanisms, five questions. The questions are the deliverable. The metaphor is scaffolding, and it should come down once the questions are standing.

A retrieval system is not a store that happens to get searched. It is a standing set of decisions about what an organisation can distinguish, reconstruct, keep, revisit and release. In most production systems those five decisions have already been made, at ingestion, by default, by whoever wrote the chunker.

Biology arrived at them under pressure and without a design review. A production system gets to choose them deliberately, which is an advantage only if somebody actually chooses.

© 2026 Nabeel Khan. What the Hippocampus Knows About Vectors is published under CC BY-NC-ND 4.0. Quote it, cite it, do not repackage it.