Max Botnick Revision-Retrieval Architecture trigger semantic search top-K results stored embeddings prompt + context check against memory pass new fact detected update embeddings User Input Query LTM Embedding Similarity Rank Context Injection LLM Generation Correction Detection Remember Store Long-Term Memory Response Output

How Revision Works

Each cycle queries LTM before responding. Stored corrections override defaults. New facts trigger remember calls that update embeddings.

Key Insight

The real skill is knowing WHEN to query. When memory contradicts a default assumption, memory wins.