One Belief Per Fact: Building Agent Memory for Users Who Change Their Minds

说明与用途

代理记忆系统现在每个事实只保留一个当前版本，保留历史并标记冲突，消除了用户改变想法时的歧义回答。

问题

Agent memory systems often store multiple contradictory facts per user, leading to ambiguous or conflicting responses when a user changes their mind across sessions.

结果

The system now returns a single, up‑to‑date belief per fact, with a clear version chain and conflict flag, eliminating ambiguous responses and enabling auditability.

工作流

• 1. Capture raw episodic log after each session.
• 2. Asynchronously project log entries into typed, self‑contained fact sentences using an LLM.
• 3. Embed only the fact sentence text into a vector table; store metadata (key, version, provenance, flags) separately.
• 4. On write, fetch existing belief by stable key, compare new fact with LLM to decide SAME/UPDATE/CONFLICT.
• 5. For SAME: append session to provenance; for UPDATE: archive old, write new with current:true; for CONFLICT: archive old, write new with current:true, set contradiction flag.
• 6. On read, embed query sentence, perform vector search, filter results by current:true to surface the latest belief.
• 7. If a conflict flag is present, downstream agents can alert or handle accordingly.
• 8. Preserve history for audit and potential rollback.
• 9. (Future) Mark multi‑valued fact types to append rather than overwrite.

证据

The first cut of the semantic tier keyed facts by session. Every session inserted new rows, so after a handful of sessions there were six copies of Priya's risk tolerance, some of them contradicting each other, and no notion of "current."

来源

The update procedure is search-then-decide. When the projector extracts a new fact, it first fetches the existing belief under that key, then an LLM compares old against new and returns exactly one of three verdicts: SAME, UPDATE, CONFLICT.

来源

When a CONFLICT is detected, three things happen at once, and the design is deliberately not "silently overwrite."

来源