Same agent-memory accuracy with 8x fewer tokens by changing the embedding model (github.com)

Turnbook has announced a significant update to its memory management capabilities, achieving a memory accuracy of 91.8% on the LongMemEval benchmark with just 3.1k tokens, which is eight times fewer tokens than previous iterations. This new version, 0.10.0, utilizes the EmbeddingGemma model for search, eliminating the need for costly model tokens to save memory. Users can now maintain a persistent, context-aware memory across different AI agents on their local machines without relying on external cloud services, thanks to Turnbook's architecture that stores every conversational turn verbatim in a single SQLite file. This development is crucial for the AI/ML community as it enhances the efficiency of AI systems in managing conversational context without incurring high token usage or costs. With improved retrieval methods that integrate word meaning, timing, and even multi-language capabilities, Turnbook allows various AI interfaces like Claude Code and Codex to interact seamlessly with the same memory pool. This seamless orchestration paves the way for future applications, where AI can better assist users by recalling past interactions, making it a game-changer in AI-driven conversations and multi-agent collaboration.