Engram – persistent memory for AI agents, in-process, no cloud (github.com)

Engram, a new in-process memory solution for AI agents, has been launched, enabling these agents to hold persistent memory without relying on cloud infrastructure or complex setups. This innovation allows developers to integrate long-term memory into their AI tools seamlessly—just a two-minute installation with the command `npm install -g @hbarefoot/engram` can give an AI agent the memory capabilities akin to a seasoned colleague. Engram operates locally using SQLite for storage and bundled embeddings, ensuring that no data leaves the user's machine, thereby enhancing privacy and security. For the AI/ML community, the significance of Engram lies in its simplicity, efficiency, and commitment to user safety. It features automatic secret detection to prevent sensitive information from being stored, and it offers robust functionality including memory feedback loops, conflict detection, and deduplication of memories. With negligible latency during recall (under 200 ms for cold starts and approximately 4 ms for warm recalls), Engram aims to sharpen AI memory over time as it's used. This local-first approach, which contrasts sharply with other memory systems that typically require cloud setups and substantial infrastructure, positions Engram as a pioneering tool that balances performance with user-friendly design.