LLMSlim gives AI agents and MCP clients long-term memory that runs on your own machine. Add it to Claude Code with one line (claude mcp add llmslim -- llmslim mcp), or to Claude Desktop, Cursor and any MCP client with one JSON entry.
It has five tools: remember, recall, list_memories, forget and fit_context. Messages are stored word for word in local SQLite, or your own Postgres for a team, and recall returns only the ones that answer the question, within a token budget you set. There is no LLM call inside, so writes cost nothing and nothing gets rewritten into a wrong fact.
The user and agent scope is fixed when the server starts, so injected text cannot read another user's memory. fit_context trims a huge log or document to a budget before the model reads it, keeping the lines that answer the query, in order.
MIT licensed, 7,000+ PyPI downloads, listed on the official MCP Registry and rated A on Glama. It also ships LangChain, LlamaIndex, Haystack and Pydantic AI adapters and a Mem0-compatible API.
No comments or reviews, maybe you want to be first?