LycheeMem

LycheeMem is a compact memory framework for LLM agents. It starts with efficient conversational memory, through structured organisation, lightweight consolidation and adaptive retrieval, and aims to extend toward action-aware memory. The README says memory is split into three stores: Working Memory (session turns, summaries and token budget management), Semantic Memory (a typed action store) and Procedural Memory (skills).

Semantic memory is built from typed, action-annotated records, with 7 MemoryRecord types including `tool_affordance` for what a tool can do and when to use it. The storage layer is SQLite with FTS5 full-text search plus a LanceDB vector index; a March 2026 news item says this "compact" design replaced a Neo4j dependency. Before retrieval, an action-aware planner turns the query into a `SearchPlan` with semantic queries, pragmatic queries about actions and constraints, tool hints and an optional time filter, and results are combined across channels including a temporal one. The README also mentions record fusion and usage statistics intended for RL.

You can use it through a web demo, an API, an MCP server or an OpenClaw plugin. Setup needs Python 3.11+ and an LLM API key (OpenAI, Gemini or anything litellm supports): `pip install -e ".[dev]"`, then configure `.env` with the model, API key and an embedding model. The licence is Apache 2.0, the project went open source on 2026-03-23, and the last push was 2026-08-06. This review is based on the README and repository data; we have not run it ourselves.