research
Half the Needed Memories Were Never Given a Chance
Causal Memory Policy intervened on retrieval itself after store-level tests left memory utility unidentified.
Summary
Causal Memory Policy intervened on retrieval itself after store-level tests left memory utility unidentified.
Memory systems often estimate whether a stored item matters only after the retriever surfaces it. Causal Memory Policy shows that this creates a positivity failure: store-level interventions cannot identify the value of a memory that is never retrieved. In the authors' tests, utility was unidentified for 54 percent of required LongMemEval memories and 67 percent on LoCoMo, including in a deployed system. Reserving context slots for memories sampled with known propensities raised required-versus-nonrequired discrimination from 0.54 to 0.66 AUC. Even then, utility estimated for one query did not predict value on unseen queries, limiting its use for irreversible retention decisions.
Why it matters
Causal Memory Policy intervened on retrieval itself after store-level tests left memory utility unidentified.
Limits and context
- Memory systems often estimate whether a stored item matters only after the retriever surfaces it.
- Causal Memory Policy shows that this creates a positivity failure: store-level interventions cannot identify the value of a memory that is never retrieved.
- Even then, utility estimated for one query did not predict value on unseen queries, limiting its use for irreversible retention decisions.
Key claims
Causal Memory Policy intervened on retrieval itself after store-level tests left memory utility unidentified.
Qualification: Memory systems often estimate whether a stored item matters only after the retriever surfaces it.
Evidence: source-2026-10-02-001
Sources
- arXiv preprint 2610.02070arXiv · primary research
Corrections
No corrections have been recorded for this story.