The standard RAG pattern treats all stored content as equally valid, but real information evolves and contradicts itself. “I love my job” (Week 1) and “I’m thinking about quitting” (Week 2) both retrieve with equal confidence because embeddings have no concept of recency, supersession, or conflict resolution. Without explicit resolution logic, the agent fails to recognize that the newer fact invalidates the older one.
This is a deeper version of what Similarity is not relevance — relevance requires reasoning identifies in search: similarity is not truth, and truth requires temporal reasoning that embeddings fundamentally cannot perform. The fix requires explicit conflict resolution logic — archiving old facts as “past history” when contradicted by newer ones — which is why The three-layer AI stack: Memory, Search, Reasoning places a Reasoning layer on top of Search. Without that reasoning layer, memory systems “remember everything but know nothing.” The practical fix is Evolving summaries beat append-only memory — rewrite profiles, don't accumulate facts — rewriting category profiles rather than accumulating facts sidesteps the temporal problem entirely, ensuring summaries always reflect current truth.