ThinkfleetAI/memmesh

What it solves

MemMesh provides a persistent, local-first memory system for AI agents, eliminating the need for hosted vector databases, external search clusters, and mandatory LLM calls for memory extraction. It prevents agents from forgetting information between sessions while keeping data entirely on the user's machine for privacy and cost-efficiency.

How it works

It is delivered as a single Rust binary that uses a local database (SQLite or Postgres) and a small local embedding model (bge-small) to enable semantic and keyword hybrid search. It integrates with AI tools via the Model Context Protocol (MCP), allowing it to be wired into tools like Claude Code, Cursor, and Windsurf. The system can automatically capture substantive information from prompts using heuristic filters to drop conversational filler without requiring an LLM.

Who it’s for

Developers and users of AI agents (such as those using Claude Code, Cursor, or Windsurf) who want their agents to have durable, private, local memory without managing complex infrastructure or paying for API calls.

Highlights

• Local-First Architecture: One binary and one file; no external vector DB or mandatory API calls.
• MCP-Native: Seamlessly integrates with popular AI coding tools and agents.
• Hybrid Search: Combines semantic search (via local embeddings) and keyword search.
• Bi-temporal Time Model: Tracks both when an event happened and when it was learned.
• Automatic Capture: Uses heuristic filters to observe and store substantive facts and preferences automatically.
• Heuristic Extraction: Captures information locally without needing an LLM for extraction.