Source: Guo, Wang, Zhu et al., "MemCredit: From Relevance to Utility for Decision-Time Memory in Long-Horizon Web Agents", arXiv 2611.15600, 11 Nov 2026. Surfaced by Elvis Saravia and DAIR.AI.
The claim
Most agent memory systems retrieve the memories most semantically similar to the current state. The paper's counterfactual analysis finds that memory utility is sparse and strongly state-dependent: most retrieved memories do not change the agent's next action, and similarity is a poor predictor of which ones do.
How it works
Measure credit offline. Run controlled interventions: replay a decision with and without a given memory in context and record whether and how the action changes. That difference is the memory's credit for that state, reference context and policy.
Distill an estimator. Train a lightweight model to predict a memory's credit from features of the state and the memory.
Select at inference. Semantic retrieval still produces a high-recall candidate set; the utility estimator decides which candidates fit in a fixed token budget. The agent policy is unchanged and no extra environment interaction is needed.
Results (author-reported)
ALFWorld, 500-token memory budget: 86.57% success versus 65.67% for BM25 retrieval.
About 152 memory tokens per decision, roughly 75% fewer than BM25 (per the DAIR summary).
Also evaluated on WebShop.
Why it matters
Memory is a context-budget problem. Filling the budget with similar-but-useless items costs tokens and can distract the policy.
The pattern generalizes: use cheap retrieval for recall, then a learned or rule-based gate for precision against a budget.
The intervention method is also a diagnostic. Even without the estimator, ablating memories on logged decisions tells you which memory types ever matter.
Tradeoffs and failure modes
Credits are policy-conditional. Swap the model and the estimator may need retraining.
Offline interventions cost compute and need replayable decisions; production agents with side effects are harder to replay.
Benchmarks are ALFWorld and WebShop, both short-action-space environments; coding or research agents may behave differently.
A utility gate can starve rare-but-critical memories (a constraint mentioned once) if they were rarely decisive in training data.
Review checklist
Do we know what fraction of retrieved memories ever changes an action?
Is retrieval tuned for recall, with a separate precision step under a token budget?
Can we replay decisions to ablate memories offline?
Are hard constraints kept out of the utility-ranked pool so they are always loaded?
Related entries
R040 (evidence-linked agent memory), R059 (the memory tax), R083 (agent memory pattern catalog), R084 (what an always-on agent should store), R002.