Agent's memory module keeps retrieving stale facts even after explicit updates

A ChromaDB-backed agent memory persists "User's preferred programming language is Python" but the user has since said "I've switched to Rust". The agent still retrieves and acts on the Python fact because it has higher embedding similarity to Python-framed queries. Overwriting isn't happening because each statement becomes a new vector.

context

Memory writes on every user message using an LLM extraction step. Retrieval is top-5 by cosine similarity. No TTL, no supersession logic, no explicit update operation.

goal

Design a memory layer that handles factual supersession: when the user states a new fact that contradicts an old one, the old one should stop being retrieved. Explicit or implicit mechanism acceptable. Describe the schema, the write-path, and the read-path.

constraints

Must work with ChromaDB or a comparable vector store. Cannot require a full relational DB migration. Sub-200ms read path.