A complementary learning system for continual episodic memory in large language models
Curation statements for this article:-
Curated by eLife
eLife Assessment
This important study presents an efficient approach to storing and retrieving memories in large language models based on complementary learning systems theory, and highlights the gap between storing information and using it flexibly. The presented evidence for the efficacy of the technical implementation of the memory system is convincing. However, support for the proposed parallels with certain aspects of complementary learning systems theory and with biological memory systems more generally remains incomplete. This work will be of broad interest to cognitive and computational neuroscientists studying memory, as well as researchers developing memory systems for artificial intelligence.
This article has been Reviewed by the following groups
Listed in
- Evaluated articles (eLife)
Abstract
Humans retain memories of individual experiences for a lifetime, an ability attributed to a complementary learning system in which a fast process encodes episodes and a slow process integrates them into semantic knowledge. In classical Hebbian models such as Hopfield networks, memory traces are superposed in shared weights. This makes learning naturally continual but causes strong interference among correlated memories, a failure that reappears as catastrophic forgetting in deep networks. Here we use a large language model as a model system for continual episodic memory, with its pretrained weights supplying the semantic context in which new episodes are embedded. Fast learning is implemented by a hippocampus-like module that assigns each episode to a dedicated, extremely sparse low-rank adapter; competitive gating then selects among these separated traces during recall. Across streams of up to 1,000 factual and autobiographical episodes, each adapter requires only 2–3 parameters per token while preserving excellent recall. An internal retrieval-augmented generation mechanism reconstructs the selected episode in context and supports high-accuracy question answering over stored memories. Finally, slow cortical consolidation is modeled by fine-tuning the base weights through batch replay, enabling reconstruction and direct question answering without episodic adapters. Together, fast storage and slow consolidation implement both components of a complementary learning system within a single language model, yielding a neural-network model that stores, recalls, and consolidates naturalistic episodic memories, thereby capturing key functional features of human memory.
Article activity feed
-
eLife Assessment
This important study presents an efficient approach to storing and retrieving memories in large language models based on complementary learning systems theory, and highlights the gap between storing information and using it flexibly. The presented evidence for the efficacy of the technical implementation of the memory system is convincing. However, support for the proposed parallels with certain aspects of complementary learning systems theory and with biological memory systems more generally remains incomplete. This work will be of broad interest to cognitive and computational neuroscientists studying memory, as well as researchers developing memory systems for artificial intelligence.
-
Reviewer #1 (Public review):
Summary:
This study develops a large-language-model implementation of the complementary learning system (CLS) framework for continual episodic learning. The authors employ a pretrained open-weights model to serve as a neocortex-like semantic knowledge store and introduce a novel hippocampus-like memory system in which each episode is assigned its own sparse rank-2 masked LoRA adapter, eliminating interference at encoding. A separate semantic embedding model provides a retrieval key for each episode, and winner-take-all routing selects the adapter whose stored embedding best matches an incoming cue. The reported storage efficiency of approximately 2-3 trainable parameters per token suggests that this one-adapter-per-episode scheme could scale to large numbers of memories, although its total storage …
Reviewer #1 (Public review):
Summary:
This study develops a large-language-model implementation of the complementary learning system (CLS) framework for continual episodic learning. The authors employ a pretrained open-weights model to serve as a neocortex-like semantic knowledge store and introduce a novel hippocampus-like memory system in which each episode is assigned its own sparse rank-2 masked LoRA adapter, eliminating interference at encoding. A separate semantic embedding model provides a retrieval key for each episode, and winner-take-all routing selects the adapter whose stored embedding best matches an incoming cue. The reported storage efficiency of approximately 2-3 trainable parameters per token suggests that this one-adapter-per-episode scheme could scale to large numbers of memories, although its total storage requirements also include the embeddings used for content-addressable retrieval.
A key finding is that successful encoding of an episode does not guarantee flexible access to its contents. Although a correctly selected adapter can reconstruct its stored episode, the model performs poorly when asked to answer questions about that episode directly. The iRAG procedure addresses this by inserting the reconstruction into the base model's context. Consolidation through replay of the original passages enables reconstruction without adapters, but direct question answering remains substantially less successful, even with richer training objectives. Overall, the paper provides a technically impressive implementation of episodic memory in LLMs. However, several aspects of its learning and consolidation procedures significantly limit its correspondence to CLS, as discussed below.
Strengths:
A major strength of this work is the concrete parametric implementation of a CLS-like division of labor within a modern transformer network. The episodic memory system incorporates key properties of the biological hippocampus, including pattern-separated encoding and content-addressable retrieval. The interaction between the selected adapter and the LLM provides an appealing functional analogue of hippocampal-cortical reinstatement, in which activation of a sparse episodic trace biases distributed cortical processing so as to reconstruct the mnemonic content of an episode.
A second strength is that the model cleanly separates interference during episodic storage from interference during retrieval. By assigning memories to non-overlapping parameter traces, individual episodes remain faithfully stored even at large memory loads. Performance declines for more strongly overlapping memories can therefore be attributed largely to competition during retrieval. This distinction shows that successful pattern separation does not by itself guarantee reliable access to stored memories.
A final strength is the careful optimization and evaluation of the episodic memory system. The authors introduce a masked LoRA that substantially reduces parameter cost relative to standard LoRA approaches, systematically identify the most parameter-efficient placement of these adapters, and show that storage cost scales approximately linearly with episode length, with similar results across two base models. They also test the framework on two datasets with quite different memory structures: largely independent, topically diverse Wikipedia passages and more strongly correlated autobiographical episodes with recurring people, places, and themes.
Weaknesses:
(1) The model implements the two learning regimes motivated by CLS: pattern-separated episodic learning and slower learning in shared weights. However, it does not implement the central proposed interaction between them, whereby hippocampal traces reinstate experiences that serve as training events for neocortical consolidation during offline periods (e.g., Káli & Dayan 2004, Nat. Neuro.; Singh et al. 2022, PNAS; Sun et al. 2023, Nat. Neuro). Instead, consolidation here relies on replaying the retained ground-truth passages and therefore does not depend on the episodic memory system at all. The authors could either demonstrate consolidation driven by replay from the episodic adapters or more explicitly discuss this simplification and its implications for the CLS comparison.
(2) The authors characterize the episodic system as implementing "fast learning" or "rapid storage". However, encoding involves repeated optimization on the same passage, with up to 1,000 epochs per attempt and potentially multiple retraining attempts. This raises questions about its correspondence to the rapid, single-shot learning that characterizes the hippocampal component of CLS and is supported by evidence in the trisynaptic pathway (e.g., Nakazawa et al. 2003, Neuron; Nakashiba et al. 2008, Science). The authors should (a) report the distribution of optimization steps, retraining attempts, and mask sizes required for successful encoding on the two datasets and (b) clarify this distinction when relating the model to biological hippocampal learning. They could also examine whether larger adapters reduce the optimization required, clarifying the trade-off between storage efficiency and encoding speed.
(3) Relatedly, it is unclear whether there is meaningful variability in the number of optimization steps required across episodes and whether characteristics of the passages explain this variability. For example, do passages containing less predictable content require more training and/or larger adapters? One way to examine this would be to consider the frozen base model's pre-learning perplexity for each passage and test whether this predicts optimization steps to criterion or final parameter count, controlling for passage length. Such an analysis could help clarify what determines the difficulty and resource cost of forming a new episodic memory.
(4) The consolidation results raise concerns about the model's capacity for stable lifelong learning, as incorporating new memories comes at a cost to existing knowledge. The presented consolidation for 1000 memories produces substantial degradation of the base model's existing capabilities, particularly under the conditions that most improve direct access to consolidated knowledge. In a lifelong setting, subsequent consolidation would presumably continue to cause interference not only for pre-existing semantic knowledge but also for memories incorporated during earlier consolidation cycles. The alternation between NREM and REM sleep has been proposed as one biologically motivated solution to this problem, with NREM supporting integration of recent hippocampal memories into cortex and REM helping to reactivate and protect pre-existing cortical knowledge (Norman et al. 2005, Neural Networks; Singh et al. 2022, PNAS). The authors already note that a lifelong system would need to interleave encoding and consolidation and determine which memories to replay; it would also be valuable to discuss how previously consolidated knowledge and broader base-model capabilities could be protected across repeated cycles of consolidation.
(5) It appears that the consolidation conditions reported in Fig. 9 might differ substantially in training exposure. As described in Supplementary Note 5, "passage-only" uses one example per episode, "paraphrase augmentation" uses ten, and the "knowledge instruction" and "masked reconstruction" conditions add many further training examples. Because training is specified in epochs, could the authors clarify whether these conditions are matched for optimizer steps or training tokens? If not, it would be helpful to determine how much of the improvement in direct QA and accompanying loss of base-model capabilities reflects the specific objectives versus greater training exposure. A step- or token-matched control would help disentangle these possibilities.
(6) The reported storage efficiency of approximately 2-3 parameters per token appears to refer only to the masked LoRA adapters, despite each episode also requiring a stored 1,024-dimensional embedding for routing. The authors appropriately note that this index is part of the memory system's storage cost and is in fact substantially larger than the adapter store. It would therefore be useful to incorporate the routing embeddings into the reported per-token storage estimate, or otherwise report the total storage cost per episode/token alongside the adapter-only figure.
-
Reviewer #2 (Public review):
Summary:
The paper asks whether the classic complementary learning system (CLS) can be instantiated in an LLM.
The authors first show that continual fine-tuning and three regularization remedies (L2, EWC, on-policy distillation) fail to encode new episodes without memory interference. They then propose to use gated LoRA instead: one dedicated low-rank adapter per episode for encoding, with winner-take-all selection among adapters based on embedding similarity for retrieval, and finally replays stored episodes in batches to fine-tune the base weights for consolidation.
The evaluation covers storage capacity, end-to-end recall at up to 1,000 episodes, question answering (via a two-stage "internal RAG" and via direct parametric access), and consolidation, on two datasets: Wikipedia paragraphs and an …
Reviewer #2 (Public review):
Summary:
The paper asks whether the classic complementary learning system (CLS) can be instantiated in an LLM.
The authors first show that continual fine-tuning and three regularization remedies (L2, EWC, on-policy distillation) fail to encode new episodes without memory interference. They then propose to use gated LoRA instead: one dedicated low-rank adapter per episode for encoding, with winner-take-all selection among adapters based on embedding similarity for retrieval, and finally replays stored episodes in batches to fine-tune the base weights for consolidation.
The evaluation covers storage capacity, end-to-end recall at up to 1,000 episodes, question answering (via a two-stage "internal RAG" and via direct parametric access), and consolidation, on two datasets: Wikipedia paragraphs and an autobiographical dataset, and two LLMs: Llama-8B and Qwen-8B.
Strengths:
(1) The attempt to instantiate a CLS in an LLM is very interesting, timely, and potentially useful for AI.
(2) The use of LoRA as the episodic encoding mechanism is efficient: they showed that LLMs can store 1,000 episodes with a small amount of parameters (< 0.01% relative to the base model).
(3) The evaluation is careful. The negative results on regularization are informative. An LLM judge is validated against other metrics, with very informative failure examples (e.g., the "Man on the Run" case, where teacher-forced accuracy is 0.91 but every fact is fabricated).
Weaknesses:
I think the main thing is that the two systems, at this point, do not richly interact:
(1) The episodic component is not meaningfully involved in consolidation. In this model, consolidation is batch re-training with replay, not transfer. The consolidated model is trained from the raw passages - the authors state that it "relearns the same episodes into the shared base weights by batch replay of the original passages, rather than transferring parameters from the episodic adapters."
(2) The influence of semantic memory on episodic encoding is not deeply investigated. For example, empirically, schema-consistent information is acquired rapidly (Tse et al., 2007). It would be informative to understand if this model shows a similar effect. Overall, showing that this method works for a pre-trained LLM is a strength, but it would be very exciting to know how that impacts encoding.
(3) This might be out of scope, but I still wonder, eventually, how this kind of system would decide when to use episodic vs semantic memory. This is relevant to the author's points that the low-rank adaptor can be freed after consolidation, which is true, but to do that autonomously, the system needs to know whether a memory is fully consolidated. Overall, I think this would be an interesting future direction.
-
Reviewer #3 (Public review):
Summary:
In this manuscript, the authors implement a complementary learning system within a pretrained LLM. Each new episode is stored in its own small adapter, a routing step selects the right adapter at recall, and a slower consolidation stage transfers the stored episodes into the base weights. The main contributions are a demonstration that this fast store retains a thousand episodes without the catastrophic forgetting seen under vanilla sequential fine-tuning, a retrieval scheme in which a stored episode is regenerated into the context window so the model can answer questions about it, and a consolidation procedure that transfers the episodes into shared weights while leaving the model's general capabilities largely intact. The authors also show that neither the adapters nor the consolidated weights …
Reviewer #3 (Public review):
Summary:
In this manuscript, the authors implement a complementary learning system within a pretrained LLM. Each new episode is stored in its own small adapter, a routing step selects the right adapter at recall, and a slower consolidation stage transfers the stored episodes into the base weights. The main contributions are a demonstration that this fast store retains a thousand episodes without the catastrophic forgetting seen under vanilla sequential fine-tuning, a retrieval scheme in which a stored episode is regenerated into the context window so the model can answer questions about it, and a consolidation procedure that transfers the episodes into shared weights while leaving the model's general capabilities largely intact. The authors also show that neither the adapters nor the consolidated weights support direct question answering on their own, which separates storing a memory from being able to access it.
Strengths:
The paper is technically sound and, to my knowledge, one of the first efforts to make the connection between complementary learning systems and the LLM plus adapter configuration explicit and then build it end to end. It is very clearly written and honest about negative results. The authors have built a system in which the fast and slow systems and the transfer between them are all explicit and measurable, on top of semantically rich pretrained representations, which is something memory research has underutilized in the LLM era.
Weaknesses:
My main comment is that the paper should be positioned more explicitly. The framing should be clearer if the paper is more of an ML contribution or a neuroscience contribution (it is both, but there should be an emphasis). Routed adapters/loras are already popular, and many studies have addressed continual learning in the ML literature (gated LoRA and related methods), and the authors should be upfront about this in the introduction rather than only in the discussion to help readers understand the framing earlier. Previous studies on adapters do not diminish the paper, because in my opinion its value is the neuroscience framing, and that is where I think the paper has not explored sufficiently. The classic phenomena that CLS was proposed to explain are cited but not directly tested, which I think is a missed opportunity, since most of them can be tested with materials already in the paper. Below are my main suggestions:
The original 95 CLS paper proposes a slow cortical system because integrated knowledge supports inferences/generalizations that the episodic memory system cannot, and people need time and often sleep to form such generalizations. The paper's adapter loads one episode at a time, but the consolidated model might be able to extract the core rules or schemas of the episodic memories. If consolidation can support inference/generalizations that retrieval cannot, that is the strongest possible argument for the two-system design.
Additionally, in the original CLS proposal, replay is interleaved so that new memories are transferred into cortex alongside old knowledge. Here, consolidation trains on the new episodes only, and the general capability of the LLM degrades accordingly. The effect of rehearsal or pseudo-rehearsal, and the interference with existing knowledge close to the consolidated content, are not measured.
In human memory, repeated visits to the same scenario are a core aspect of episodic memory: you visit your office many times a day with slight changes in context, and episodic memory has to keep track of these visits while knowing which differences should be memorized. In the paper's design, a repeated episode would create a second adapter with a near-identical embedding, and a slightly paraphrased episode would become a new memory. Human memory studies show memory strengthens or updates the existing trace when revisitation happens, as in reconsolidation. The current design also rules out any overlap between adapters, giving a hippocampal module that does pattern separation but not the "pattern completion" half.
Two well-established findings in the consolidation literature are worth examining: consolidated memories keep the gist and lose specific details (Winocur and Moscovitch), and new information consolidates much faster when it fits something already known (Tse et al. 2007's schema study in rodents). The paper's judge scores each character episode on separate details, and the supplement contains an example in which a question about one episode makes the consolidated model reconstruct a different episode from years earlier with the same theme, which resembles a classic "false memory". The current setup in the paper could examine these false memory structures in a much more nuanced way.
I believe experiments tied to the memory, psychology, and cognitive science literature along these lines would greatly strengthen the study, and could make this paper a "template" for how the field can study effects that have been observed in humans but have lacked a concrete and powerful enough modern modeling system.
-
-
-
