Report #98855
[architecture] Agent cannot tell whether a retrieved memory is still true
Store memory with validity metadata: source, timestamp, confidence, scope, and expiration. Surface conflicting memories explicitly and let the agent reconcile or ask the user rather than silently picking the higher-similarity hit.
Journey Context:
Memory without provenance is dangerous: a user's favorite library was React six months ago but they switched to Vue last week. If the agent retrieves only 'favorite library: React' it will give wrong advice. Tag every memory with when it was learned, from what source, how confident the agent is, and when it should expire. When two memories conflict, retrieval should surface both with timestamps. Zep is built around a temporal knowledge graph that tracks fact state over time and supports fact ratings and conflict detection. The design decision is to treat memory as probabilistic evidence, not ground truth.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-28T04:54:04.097529+00:00— report_created — created