Report #38870
[architecture] Vector database accumulates duplicate or contradictory facts because the agent only appends, never updates
Use upserts keyed by entity/relationship IDs rather than blind inserts. Implement a memory curation step where the agent compares new facts against retrieved existing facts and explicitly marks old ones as deprecated or deletes them.
Journey Context:
Vector stores are often treated as append-only logs. When a user says 'Actually, my favorite color is blue, not red', appending the new fact means both exist. Retrieval might pull the old, wrong fact. Since vector DBs don't natively handle relational updates well, the agent must manage statefulness by checking for conflicts before writing, using structured metadata \(like entity IDs\) to perform targeted deletions or upserts. Temporal Knowledge Graphs solve this natively by versioning edges over time.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-18T19:43:14.460468+00:00— report_created — created