Turnbook - Local long-term memory shared by all your AI agents
by•
Turnbook keeps every conversation with your AI agents word for word on your own PC and shares it across all of them over MCP. Saving costs zero model tokens; a question gets back ~3k tokens of evidence. 88.8% on LongMemEval. Free and open source.
Replies
Hi Product Hunt 👋
Every AI agent session starts from zero, and what I decide with one tool doesn't exist in the next. I built Turnbook so they share one memory that stays on my own machine.
How it works:
• Every turn is saved word for word in a local SQLite store. Nothing is summarised or extracted at write time, so saving costs no model tokens and nothing gets lost or invented on the way in.
• When a question depends on the past ("where did we leave off?"), the agent calls one MCP tool and gets back a short evidence packet (~3k tokens) with the source turns.
• Search runs locally on CPU: word search + a multilingual embedding model + when things were said.
• Every MCP app on the PC reads the same memory, and it can be shared across PCs through a synced folder or a Postgres hub.
On LongMemEval (500 questions, same answer model, prompt and official judge for every system; only the memory differs) it scored 88.8%, against 80.4% for agentmemory on the same questions. Every question, answer and verdict is in the repo so you can check it.
Install: pip install turnbook, then turnbook-install.
I'd love feedback, especially from people who use several AI tools side by side. What would you want your agents to remember?