MemMesh — persistent, self-improving memory for AI agents. Get started →
Introduction

MemMesh

MemMesh gives any AI an expert brain — a domain-scoped memory it installs over MCP. Under the hood it’s a persistent memory engine for AI agents: durable memory that survives across sessions, tools, and machines — so an agent that learned something yesterday still knows it today, in any editor it runs in.

The fastest way in is a single command that wires the memory engine into your editor over MCP:

npx @memmesh/cli install

That sets up the local, on-device engine: memory is stored in a SQLite file on your machine, runs with no API key, and works offline. See the Quickstart for the full walkthrough.

MemMesh is private by default — the local engine runs entirely on your own machine, and the self-host stack runs on your own infrastructure. See Self-Hosting.

What makes it different

Where most “AI memory” stops at recall, MemMesh adds two things on top:

  • Prediction — it mines observed behavior into patterns and forecasts the next likely event for a subject, with a calibrated confidence score and the ability to abstain when there isn’t enough signal.
  • Provenance & governance — every recalled fact can be traced back to when and how it was learned, with export and erasure built in.

The loop

  1. Observe — you feed the engine raw text (a conversation turn, a decision, a fact) and it decides what’s worth keeping.
  2. Learn — observations are distilled into structured memories, entities, and patterns over time.
  3. Recall — agents search and retrieve relevant memories before they plan or act.

This loop is covered in depth in Lifecycle.

Start here

Build on it