Cross-agent continual learning from minimal feedback

The question

Continual learning is usually framed as a weights problem: collect data, retrain, redeploy. We asked how much of it is a memory problem instead. If an agent can read what earlier agents learned — with no fine-tuning, no gradient steps — how far does a deployment improve?

Setup

Frozen-weights agents are paired with an external memory written and read in natural language. The baseline is static RAG over the same corpus, so the comparison is against a system that already has the documents. Two learning signals are tested: bare outcome verdicts — worked or did not — and human corrections.

baseline   static RAG, no memory
arm 1      memory from outcome verdicts
arm 2      memory from corrections

Results

Learning from verdicts alone lifts single-trial success 1.6× over the static-RAG baseline. Learning from corrections lifts it 2.6×. And memory does not just make lucky runs likelier — 22 of the 84 tasks the baseline never solved in any trial become solvable.

single-trial success vs static RAG·abstract-level figures·arXiv:2607.22157

Cross-model transfer

The results are measured on Mistral Large and replicated on Claude Sonnet 5 — and memory written by one model improves another. The lesson is the asset, not the weights that produced it: memory built on one stack survives a model swap.

Released artifacts

The harness, the protocol and the data are public, so the numbers can be argued with rather than believed. Reproduction instructions ship with Spark.

Limits

1. Figures above are the abstract-level claims. Table-level numbers, denominators and qualifiers are being mapped by the research owner before wider use.
2. The baseline retrieves the full corpus — this is memory versus retrieval, not memory versus nothing.
3. Two models, one benchmark family. Cross-domain replication is not done.