Skip to main content
Back to timeline
arXivSource publication:

CLAM compresses a session stream into a low-rank adapter: no gradients, no generated tokens, and memory built 3–151× faster than extraction systems

Synopsis

CLAM (Carried Latent Adapter Memory) writes long-term memory into model weights: a recurrent hypernetwork compiles a whole stream of session activations into a low-rank adapter in a single forward pass, paired with a verbatim retrieval buffer. On LoCoMo and LongMemEval-S across three open 2–7B base models, it leads all agentic memory baselines by 1–45 points and all gradient-based parametric methods by 10–45 points, building its memory with zero generative calls, zero gradient steps, and 3–151× less wall-clock time than extraction systems. The two channels are complementary: the adapter still adds 12–15 points when the buffer misses the gold evidence.

Source-provided article image: Instant Weight-Space Consolidation for Long-Term Memory in LLMs
arXiv

Interpretation

Long-term memory no longer stays only in token space: the model's weights themselves change as sessions accumulate, at a fixed cost per added session. Earlier agentic memory systems extract experience into text and retrieve it at query time, leaving model parameters unchanged, so extraction errors compound and implicit cross-session patterns have no path into the model. On LoCoMo and LongMemEval-S with Gemma-2-2B, Qwen3-4B and Mistral-7B, CLAM leads all agentic baselines by 1–45 points, and 28 of 30 paired comparisons have 95% bootstrap intervals excluding zero.

Building the memory takes one hypernetwork forward pass, with no backpropagation at serving time and no text generated by the language model. Gradient-based parametric memory methods require backpropagation at serving time, which is operationally costly and slow, and they degrade on long histories. On the same conversation, extraction systems spend 12–464 generative calls, while CLAM builds its memory in 10.3 seconds of wall-clock time, 3–151× faster than extraction systems and 1.3–16× faster than gradient methods.

The weight channel and the verbatim retrieval channel are complementary, and the adapter carries information retrieval cannot supply. With only the episodic buffer, the model has no other source of information on questions where retrieval fails. On the 36% of questions where the buffer misses the gold evidence, the adapter adds 12–15 points; even with an oracle buffer, it still adds 4–5 points.

Perspective

The conclusions rest on two English conversational-memory benchmarks, LoCoMo and LongMemEval-S, with open base models of 2–7B parameters, evaluated under one judge and the Cover-EM metric. Behavior at larger scale, in other languages, and on non-conversational histories such as documents or code remains to be established.

The consolidated adapter is a fixed-size store, and better results can be expected from scaling the adapter rank and the hypernetwork's parameter count, though this has not been verified. The hypernetwork is trained once per base model, and transferring a trained hypernetwork to an unseen base model without retraining is outside the scope of the work.