Recent advances in optogenetics and fluorescent protein markers have helped neuroscientists locate brain cells corresponding to individual memories (engrams). This post explains how such representations might physically and semantically shift.

Background

“Encoding” is the short-term enpatterning of neurons to store a memory. This seems to happen in the hippocampus, a small horn-shaped structure in the brain’s center.

Researchers have found that memories can be encoded in one neuron. These “engrams” were found using genetic edits which dyed cells activated during encoding. The preliminary model is that Hebbian associative learning and competition between prospect engrams is sufficient for their development; see here for visuals.

“Consolidation” occurs when the initial memory code is suffused to broader regions, e.g. the neocortex. It’s still an open question whether — and how much — consolidation affects brain maturation. By my view, that is.

Representational-hierarchical theory

The neocortex is the wrinkly surface of human brains. Like modern AIs, it’s a feedforward network. It gets input from deeper regions and outputs on the brain surface; longer neurons then wrap the results back into deeper regions.

Representational-hierarchical theory posits that this very large brain region contains lower-level representations of the world (softness, yellowness), while the hippocampus contains relatively more complex representations (dog, homotopy type, hibiscus growing conditions).

In particular, R-H postulates that structures are compositional; the hippocampal representations are composed of simpler cortical ones.

I can imagine this being very strange if you didn’t first learn about hippocampal indexing! There are more complex things than there are simple things. Why would the smaller region contain combinatorially larger structures?!

Hippocampal indexing

See Goode et al^[1] for a more formal, less speculative treatment. Their article inspired this (my first distillation!) post.

Indexing theory proposes that engrams “index” representations involved in a memory. Instead of holding the entire construct “walking with my dog in the forest”, which has a lot of information, an engram associates the preexisting constructs “walking” + “my dog” + “forest”.

Under R-H theory, engrams point to relevant neocortical areas. Memory recall activates a single compressed representation which unfurls into many atomic associations. It’s this data compression which makes indexing efficient.

Initial engrams, however, last merely weeks. How then do we maintain memories for decades?

Brief anatomy

Representations are physically and semantically shifted through the hippocampus during consolidation; in fact, they move between histologically distinct regions for long-term storage^[2].

The trisynaptic circuit is the core of the hippocampus. It starts in the dentate gyrus, which seems to finely differentiate between closely related patterns and memories^[3]. This then projects to cornu ammonis 3 (CA3), a tangled recurrent network; then finally to CA1 and the rest of the brain.

I find it interesting that the DG is, at a systems level, context-precise, while CA3 is autoassociative. As we’ll see, DG engrams seem to be narrower and correspond directly to episodic memories; CA3 engrams are broader and more general.

Consolidation as generalization

It’s conventionally believed that consolidation moves representations from the hippocampus to the cortex. This is very likely true! But it’s much harder to trace networks through a whole brain than through part of the trisynaptic circuit, so cell-specific work has been limited to within the hippocampus.

Ko et al, through excellent histological analysis, found that representations are encoded in the DG and thereafter consolidate into CA3.

They further found that CA3 representations are more general than DG. Genetically altering DG neurogenesis correspondingly augmented behavioral generalization! I’ve noticed before that despite my terrible episodic memory, I ingest and intuit topics rapidly. Perhaps I have higher than typical hippocampal neurogenesis; might this also explain my poor working memory? Alas!

It does seem that whatever construct is relevant to a memory’s recall eventually moves to the neocortex, since hippocampal damage affects memories only up to 15 years old. This is indeed an extremely long potential consolidation period which I’d love to see traced out per-cell.

Discussion

Some thing remain unclear.

How does intrahippocampal consolidation occur? I’d especially like to see something simulatable.

Are there any network-mediated ways by which sleep affects cortical consolidation? Inasmuch as engram-mediated Hebbian cortical learning exists, perhaps spontaneous engram activation is more effective in an electrically quiet environment (sleep). Might spontaneous engram activation cause dreams? Would spontaneous activation help associate structures at different representationally-hierarchical levels? Do people with higher hippocampal neurogenesis experience dreams with more abstract themes?

How relevant are place cells to consolidation? See here.

1. ^{^}

   Goode, Travis D. et al. An Integrated Index: Engrams, Place Cells, and Hippocampal Memory. Neuron, Volume 107, Issue 5, 805 - 820.

2. ^{^}

   Ko, S.Y., Rong, Y., Ramsaran, A.I. et al. Systems consolidation reorganizes hippocampal engram circuitry. Nature 643, 735–743 (2025). https://doi.org/10.1038/s41586-025-08993-1.

3. ^{^}

   Leal, S.L., Yassa, M.A. Integrating new findings and examining clinical applications of pattern separation. Nat Neurosci 21, 163–173 (2018). https://doi.org/10.1038/s41593-017-0065-1.