Agents don't need memory, they need documentation
8 hours ago
- Memory plugins rely on RAG: they extract snippets from conversations, store them in a vector database, and retrieve similar snippets on each prompt.
- This approach has fundamental flaws: similarity search lacks context and correctness, stored snippets lose surrounding context, the past is treated as truth despite code evolution, agents can't search for unknown unknowns, and the store is unauditable.
- The correct solution is not recall-based memory but documentation-based memory: agents should write and read structured documents (instructions, specs, decisions) instead of relying on fragmented RAG retrieval.
- Documentation-based memory changes the agent loop from 'prompt → build → forget' to 'prompt → consult → build → update'.
- The author developed Operator Memory, a system where agents use a Markdown brain folder to persist knowledge, always consulting and updating documents, without vector databases or embeddings.