Agent memory is often described as a storage problem. The harder failure appears when a stored fact used to be correct and is now wrong.
TEPA, a new preprint, models memory validity explicitly. When new evidence conflicts with an old precedent, it revokes the old record from active use while preserving the history for audit. The authors report a score of 0.950 in a complete-reversal setting, compared with 0.210 for append-only and last-write-wins baselines. Those are author-reported research results, not an independently reproduced production benchmark. The paper also reports remaining retrieval-chain and long-context limits.
The production translation can be deliberately boring. Give each durable memory a status, valid_from, superseded_by, and evidence_id. When evidence conflicts, write a revocation event, link the replacement, and record the reason. Do not erase the old record. Do not let the last write automatically become the current truth.
This changes debugging too. A failed decision can be replayed against the evidence and policy that were valid at the time. You can then tell the difference between a reasonable old decision and a system that ignored an already available correction.
Source: TEPA: Revoking Stale Memories for Conflict-Robust Language Agents.
Top comments (0)