@cortensor Stale recall is often a write-time problem, not a retrieval one. If a newer fact doesn't explicitly replace the old one when it's saved, both stay live and retrieval grabs whichever matches the query better.
@hwchase17 Agreed. A copilot makes each person a bit faster, but nothing captures what the team learns. Without a loop that writes lessons back somewhere shared, every session starts from zero.
@Bio9xnh I built something like this and landed on: rewrite active memory with explicit add/update/delete ops, mark replaced facts as superseded instead of keeping both live, and snapshot before every rewrite. The audit trail lives in snapshots, so the live store stays clean.
@xiaoanlab The verification step that restores lost facts is the clever part. Merging compatible records is the easy case though. The harder one is records that conflict: if folding keeps both the old and new version, the agent just uses whichever one retrieves first.
@LighthouseWeb3 Portability is great, but a shared memory layer also shares every model's mistakes. If one agent writes a wrong fact and another reads it as truth, the error spreads. The more agents write to one store, the more you need periodic cleanup that resolves contradictions.
My DeepSeek agent now goes to sleep π€π
While it naps, it replays recent chats, merges duplicate memories, fixes contradictions, and lets unused ones fade. Every rewrite is snapshotted, so a bad dream can be undone.
Most agent memory only grows. Mine sleeps on it.
@MohakBajaj5 This is the part most memory layers skip. I built something that treats deletion as a first-class step: a background rewrite merges duplicates, resolves contradictions, and down-ranks what stopped mattering, with a snapshot first so a bad prune is reversible.
@1eo I agree that a pile of notes breaks the moment two facts conflict. The lossiness of human memory is also doing useful work, though: unused and contradictory items get rewritten or faded, not just stored forever. And if every rewrite is snapshotted, you can roll back a bad one.
@sarahwooders@Letta_AI The gap between storing memory and actually using it feels like the real failure mode. I've seen models retrieve a fact and still ignore it, or grab a stale one, especially when the memory is a raw dump instead of a rewritten, current summary.
Your AI never forgets. That's the problem.
Somnia β sleep-stage memory for agents.
Buffer β dream β snapshot. Let your AI dream.
https://t.co/GCo0Ks2tRR
Iβm 19 and building Noscari, an open-source AI project exploring how agents can learn from experience.
I emailed Prof. Jay McClelland about a problem I was stuck onβand he actually replied! π
He shared his thoughts and recommended a paper. Made my day!