Why the brain keeps two memory systems

In May 2026 a group at Illinois ran an experiment that should have been boring. They gave an agent memory, let it consolidate what it had learned into reusable lessons, and measured whether it got better.

It got worse. Memory utility rose, then degraded, and eventually fell below the no memory baseline. Most striking of all: when the agent consolidated from ground truth solutions, correct answers it had been handed, GPT-5.4 failed on 54 percent of ARC-AGI problems it had previously solved without any memory at all.

Giving it the right answers, and letting it write down what it learned, made it forget how to do things it could already do.

That result is surprising if you think of memory as storage. It is close to predicted if you know why mammals ended up with two memory systems instead of one. This paper is about that theory, the failure it predicts, and what follows for anyone building agent memory today. It builds on the architecture we mapped in Read and write; where that paper set out the components, this one is about why the boundary between them has to be defended.

1. The problem a single memory system cannot solve

Start with the computational fact that forces the design.

In 1989 McCloskey and Cohen documented catastrophic interference in connectionist networks: train a network on a new set of items and it loses the old ones, badly and abruptly, because the same weights encode both. Learning fast means changing many weights at once, and changing many weights at once overwrites whatever those weights were already doing.

That is not a quirk of 1980s models. It is a property of distributed representations, which is what makes them efficient in the first place. Anything that stores knowledge by superimposing it across shared parameters faces the same tradeoff: fast learning destroys old knowledge, slow learning preserves it but cannot capture a single experience.

You cannot have both properties in one system. So biology built two.

2. Complementary Learning Systems

In 1995 McClelland, McNaughton and O'Reilly gave the canonical answer in Psychological Review, in a paper whose title is the whole argument: why there are complementary learning systems in the hippocampus and neocortex. Kumaran, Hassabis and McClelland updated it in 2016 with an explicit eye on what artificial agents need, which is why the theory keeps turning up in machine learning papers.

The division of labour:

The hippocampus can record tonight's dinner without disturbing your knowledge of what dinners are, because its representations are kept deliberately non overlapping. The neocortex extracts what dinners generally involve, but only across many exposures, slowly enough that no single experience distorts the structure.

The mapping onto AI systems is not a metaphor, it is close to an identity. Weights are neocortex. Slow, expensive to change, holding generalised structure, and vulnerable to exactly the interference McCloskey described, which is why fine tuning on a narrow set can quietly degrade unrelated capabilities. Context and retrieval are hippocampus. Fast, episodic, disposable, able to hold one specific thing without touching anything else.

That is the architecture almost every serious AI system has converged on, usually without anyone citing the theory. And it converged there because the same constraint was doing the same work.

3. Consolidation is slow, offline and replayed, for a reason

Here is the part the 2026 result turns on.

The brain does move information from the fast system to the slow one. That is systems consolidation, and it happens largely offline, during sleep and quiet rest, through replay: the hippocampus reactivates recent activity patterns in compressed bursts, and the neocortex, exposed to those replays repeatedly and interleaved with older material, gradually extracts structure.

Notice three properties, because they are all engineering decisions.

It is slow. Consolidation takes days to years in mammals, not milliseconds. The slowness is not a limitation of wetware, it is what prevents one experience from rewriting the schema.

It is interleaved. Replay mixes new episodes with old ones. Training only on what just happened is precisely the thing that causes catastrophic interference, so the brain does not do it.

The episode is not deleted when it is consolidated. Hippocampal traces persist alongside the neocortical abstraction, and remain retrievable. The abstraction is added, not substituted.

Now compare that to how a typical agent memory works in 2026. Something happens. An LLM immediately rewrites the trajectory into a lesson. The lesson goes into the memory bank. The trajectory is discarded. It fires after every interaction, on the newest material only, and it replaces the evidence with the summary.

That is fast, uninterleaved, and destructive. It violates all three properties, and it is the default in most agent memory systems shipping today. We should know: it is close to what we described building in our own earlier work, and it is why we are writing this.

4. What the 2026 experiment actually shows

The Illinois result deserves to be read carefully rather than as a headline.

The authors gave agents an ARC-AGI Stream environment exposing three actions: Retain, Delete and Consolidate. The findings, in their words and numbers:

• Consolidated memories produced by current LLMs are often faulty even when derived from useful experiences.
• Utility rises, then degrades, and can fall below no memory at all.
• Consolidating from ground truth solutions, GPT-5.4 failed 54 percent of problems it had previously solved unaided.
• The regression traces to the consolidation step, not the underlying experience. The same trajectories produced qualitatively different memories under different update schedules.
• An episodic only control that simply retained trajectories stayed competitive with every consolidator tested.
• Agents allowed to preserve raw episodes by default doubled the accuracy of those forced to consolidate. Disabling consolidation entirely matched the best regime.

Read against the theory, none of this is mysterious. They built a system that consolidates fast, on the newest material only, and overwrites its own source. The brain avoids all three. The failure mode is the one the architecture predicts.

The authors' recommendation lands in the same place: treat raw episodes as first class evidence and gate consolidation explicitly rather than firing it after every interaction. That is the flow above. Its one invariant is the second to last node, where whichever branch ran, the episode survived it.

There is a pleasing convergence here. When OpenAI shipped GPT-6 Astra, the Codex change they described alongside it was moving away from repeated compaction toward keeping notes across context windows with earlier context still searchable, which we noted in our reading of that launch. That is the same correction: stop letting the summary replace the record.

5. Reconsolidation, or why every read is a write

One more piece of biology, and it is the one with the sharpest engineering consequence.

In 2000, Nader, Schafe and LeDoux showed something that unsettled the field. They took a consolidated fear memory in rats, one that was days old and stable, and reactivated it. Then they blocked protein synthesis in the amygdala. The memory did not merely fail to update. It was gone.

The implication is that retrieval returns a memory to an unstable state from which it must be actively stored again. Recall is not a read. It is a read followed by a write, and in that window the memory can be altered, weakened or lost.

For anyone building a memory system this reframes the threat model. If your retrieval path can write back to its source, then every read is an opportunity for drift, and drift compounds silently because nothing errors. You only see it if you measure fidelity across many cycles, which is exactly what the 2026 study did and almost nobody does in production.

It also connects to security. ASI06 in the OWASP agentic list is memory and context poisoning, which we covered in the agent security paper. Add reconsolidation and the attack surface widens: an attacker does not need write access to your memory store if they can influence what gets retrieved and how it gets written back. A write to long term memory is a privileged operation. So, it turns out, is a read that triggers one.

6. Forgetting is a feature

The last correction is to an assumption almost everyone brings to this.

Richards and Frankland argued in Neuron in 2017 that transience is not a failure of persistence but its necessary counterpart. Memory systems that retain everything generalise worse, because outdated specifics compete with current ones and every retrieval has more interference to fight through. The goal of memory is not maximal retention, it is optimal decision making, and those come apart.

Agent memory built without an expiry policy is not more capable. It is accumulating interference, and it will degrade in a way that looks like the model getting worse. An expiry policy, a relevance decay, a rule that supersedes stale entries: these are not housekeeping. They are part of the memory system, and the theory says they always were.

7. What we would build

Six properties, following from the above rather than from taste.

1. Keep the raw episode, always. It is the evidence. A consolidated lesson that overwrites its trajectory has destroyed the only thing that could have corrected it.
2. Make consolidation an explicit gated operation, not a reflex after every interaction. The default should be retain.
3. Interleave when you consolidate. Old material alongside new. Consolidating only on what just happened is the recipe for interference, in silicon as in tissue.
4. Make retrieval read only by default. If a path can rewrite its source, that path needs a reason, a record and a test for drift across cycles.
5. Keep provenance on every abstraction. Which episodes produced this lesson. Without it a bad consolidation cannot be traced back and removed, only discovered later as inexplicable behaviour.
6. Have an expiry policy. Not because storage is expensive, but because retention is not free.

The larger point is that this is one of the places where the biology is genuinely instructive rather than decoratively so. We traced that longer history in Brain blueprints of AI. Complementary Learning Systems is thirty years old, it predicted a failure mode that a 2026 experiment then demonstrated in a frontier model, and the fix it implies is architectural rather than a better prompt. That is a theory doing real work.

If you are building agent memory and this describes your system, the audit is a short piece of work with a clear answer. Talk to us, or see how we build systems for businesses.

8. Sources

1. Zhang, Lin, Wu, Sun, Li, Li and Peng, Useful Memories Become Faulty When Continuously Updated by LLMs, arXiv 2605.12978, May 2026, revised August 2026. Source of the 54 percent result and the consolidation findings.
2. Kumaran, Hassabis and McClelland, What Learning Systems do Intelligent Agents Need? Complementary Learning Systems Theory Updated, Trends in Cognitive Sciences, 2016.
3. McClelland, McNaughton and O'Reilly, "Why there are complementary learning systems in the hippocampus and neocortex: insights from the successes and failures of connectionist models of learning and memory", Psychological Review 102(3), 1995. The original statement of the theory.
4. McCloskey and Cohen, "Catastrophic interference in connectionist networks: the sequential learning problem", Psychology of Learning and Motivation 24, 1989.
5. Nader, Schafe and LeDoux, Fear memories require protein synthesis in the amygdala for reconsolidation after retrieval, Nature 406, 2000.
6. Richards and Frankland, "The Persistence and Transience of Memory", Neuron 94(6), 2017.
7. Position: Modular Memory is the Key to Continual Learning Agents, arXiv 2603.01761, and Metric-Topology Factorization: a computational framework for hippocampal neocortical intelligence, arXiv 2603.03362, both 2026, for current work applying the theory.

Two caveats. The neuroscience summarised here is a simplification of an active field, and the hippocampal and neocortical division is cleaner in the theory than in the tissue. And the 2026 agent result is a single study on one benchmark family; it is strong evidence that reflexive consolidation can fail badly, not proof that it always will. We have cited it because it is well controlled and because the theory predicted the shape of the failure before the experiment existed.

This note sits in our Research track, alongside the retrieval and agentic threads. The full library is at Research.