LycheeMem

Why we included this project

Multi-turn agents tend to forget what earlier sessions established, and LycheeMem exists to fix exactly that. It sits between the agent and its context window, storing structured knowledge like facts, preferences, and tool-use patterns, then pulls back only the relevant fragments when they matter instead of replaying the whole conversation. The design keeps the internals visible: a working-memory manager enforces token budgets, a search coordinator runs five recall channels across SQLite FTS5 and LanceDB, and a synthesizer scores candidate fragments before assembling the context that grounds the next answer. It connects to real agent runtimes through an MCP endpoint, a native OpenClaw plugin, a Claude Code plugin, or plain Python with a FastAPI server. Teams that want to inspect and audit what got remembered will appreciate the staged pipeline and the bundled React dashboard, and because it runs on Python with SQLite plus LanceDB, it stays cheap to deploy for small agent projects.