MemPalace/mempalace

What it solves

MemPalace provides a local-first AI memory system that stores conversation history as verbatim text, avoiding the loss of detail that occurs when AI models summarize or paraphrase. It solves the problem of long-term memory for AI agents and users, allowing them to retrieve specific, exact information from past interactions without relying on cloud APIs or external services.

How it works

The system uses a structured index where people and projects are organized as "wings," topics as "rooms," and original content is stored in "drawers." This hierarchical structure allows for scoped searches rather than searching through a flat list of files.

It employs semantic search via a pluggable backend architecture. While ChromaDB is the default, it supports various other backends like Milvus, Qdrant, pgvector, and a native Rust-based engine for high-performance local search. It can be integrated into AI tools like Claude Code, Cursor IDE, and Gemini CLI via an MCP (Model Context Protocol) server, providing 45 different tools for memory management.

Who it’s for

Developers and AI agent users who want a private, local-first memory layer for their AI assistants, specifically those using tools like Claude Code or Cursor IDE who need to maintain context across sessions.

Highlights

• Verbatim Storage: Stores original text exactly as it was, ensuring no information is lost through summarization.
• Local-First Privacy: Operates entirely on the machine, with no API calls required for core retrieval.
• Pluggable Backends: Supports multiple vector stores including ChromaDB, Milvus, and a native Rust engine.
• High Retrieval Recall: Achieves 96.6% R@5 on LongMemEval benchmarks without using an LLM for retrieval.
• MCP Integration: Provides a comprehensive set of tools for AI agents to read, write, and navigate their own memory.
• Knowledge Graph: Includes a temporal entity-relationship graph backed by SQLite for tracking relationships over time.