Retrieval returns the correct record. The record is stale. And because the model states fresh and dead facts with identical confidence, nothing in the output tells you which one you got.
Your agent's memory isn't broken because it can't find the fact. It's broken because the fact it found stopped being true.
That is not a retrieval problem
It's a state-management problem — the same one distributed systems have always had: validity windows, invalidation, cache coherency. A memory layer with none of those is a cache that never expires.
What the numbers say
Memory is now a first-class layer with its own benchmarks:
- LoCoMo 92.5
- LongMemEval 94.4
- BEAM 1M 64.1
...at roughly 6,900 tokens per query on average. The biggest gains are in temporal reasoning (+29.6) and multi-hop (+23.1). And the open problems named in the same report are cross-session identity, temporal abstraction at scale — and staleness.
Stale facts get surfaced. They just don't get acted on.
On the STALE benchmark — a separate 2026 study on stale-fact handling:
- a system surfaced the updated fact about 77% of the time, but marked it as worth acting on only ~3%;
- one frontier model caught a stale fact 92% of the time when asked directly — and 30% when the question silently depended on it.
The fact wasn't missing. It was present, and discounted.
The pattern that works
Zep's Graphiti shows the shape of the fix: store facts with validity dates, mark contradicting facts invalid instead of overwriting them, and surface only the current version.
The rule
Memory is not "accumulate more." A fact needs a bounded window of validity — otherwise you're shipping an agent that is confidently wrong on a schedule.
How are you handling staleness in your agent's memory layer?
Top comments (0)