letta vs LycheeMem
letta: Letta is a powerful platform for building stateful AI agents equipped with advanced memory, enabling them to learn and self-improve over time. It provides both a command-line interface for local agent execution and a comprehensive API with Python and TypeScript SDKs for seamless integration into applications.; LycheeMem: LycheeMem is a compact memory framework for LLM agents that organizes memory into working, semantic, and procedural stores. It features a four-stage pipeline (WMManager, SearchCoordinator, SynthesizerAgent, ReasoningAgent) with a background ConsolidatorAgent, and supports multi-channel retrieval (FTS, vector, tag, temporal). It integrates via API, OpenClaw plugin, and MCP protocol.
LycheeMem is a memory framework you embed in your own agent, with working, semantic and procedural stores. Letta ships the agent itself.
TL;DR
- You want a long-lived agent that remembers past work and rewrites its own memory and context over time.
- You want one agent reachable from CLI, desktop, browser, mobile and messaging apps, with subagents, schedules and hooks.
- You want memory you can inspect: the README says context is tracked in git and can sync to your own repository.
- You only need to add memory to an agent you already use through MCP; a small MCP memory server is a lighter fit.
- You rely on older Letta tutorials or the V1 API server: the project restructured, and the V1 server is retired.
- You want everything to work without signing in: Letta Cloud is the default backend and some features need it.
- You are building your own agent and want a memory framework with working, semantic and procedural stores.
- You prefer light storage: SQLite with FTS5 plus a LanceDB vector index, with no Neo4j.
- You want typed, action-aware records, including a tool_affordance type for what a tool can do and when to use it.
- You want something stable today: the project has been open source only since March 2026, so APIs may change.
- You want zero extra model cost: memory extraction and retrieval use LLM and embedding calls.
- You expect the action-aware and reinforcement-learning parts to be finished: the README describes them as a direction.
Side-by-Side Comparison
lettaLycheeMemFeatures
lettaLycheeMemUse Cases
lettaLycheeMemBest For
lettaLycheeMemFAQ
Deploy on DigitalOcean
Host your MCP servers, Docker containers, and AI agents with $200 free credit over 60 days.