memU

memU is an open-source memory system that gives AI agents persistent, shared context across sessions, agents and devices. It distils reusable skills from an agent's own session history into inspectable Markdown, stores them in SQLite or PostgreSQL, and feeds them back automatically into Claude Code, Cursor, Codex, ChatGPT Work mode and other supported hosts.

memU is an open-source memory layer that lets a user's context follow them between coding agents, sessions and machines. Rather than each assistant keeping its own private history, memU maintains one shared, inspectable wiki of Markdown entries with source-linked recall, and exposes two operations: memorize, which slices an agent's session log into self-contained jobs and has the agent itself write reusable skill files, and retrieve, which pulls the relevant entries back into a later task automatically. It works through host adapters that run as sidecars alongside desktop agents, with published support for ChatGPT Work mode, Claude Code on both desktop and CLI, Cursor, OpenClaw, Hermes Agent, WorkBuddy and Cola, though coverage varies by operating system and the project documents the limitations per combination rather than hiding them. Storage is pluggable: SQLite as the local single-writer default, PostgreSQL with pgvector for concurrent access and larger stores, and in-memory for tests. The project deliberately keeps its core small — the README states the memory logic is roughly 500 lines, compact enough to inspect, understand and adapt — and the MemoryService performs storage, embedding and retrieval without making its own LLM calls, so the memory layer adds no hidden model spend. Written in Python and requiring 3.11 or newer, it installs via pip, npm or uv and is licensed Apache-2.0 according to the README. NevaMind AI open-sourced it on 29 July 2025 and it has since reached 14,267 GitHub stars and 1,055 forks. A free hosted memU Cloud is offered, alongside bespoke private-cloud and self-hosted deployments arranged as design-partner engagements.

Category

Agent Development

Pricing

Free, Contact for pricing

Target Market

CTOs, Enterprise Developers, AI Engineers, Platform Engineers

Deployment

Open-source, Self-hosted, Cloud-first

Founded

2025

• ✓
  Cross-agent shared memory

  One memory store follows the user between Claude Code, Cursor, Codex, OpenClaw and the other supported hosts.

• ✓
  Agent-written skill extraction

  The agent itself distils its session history into reusable Markdown skill files rather than an opaque embedding blob.

• ✓
  Inspectable Markdown storage

  Memories are plain files you can read, edit, diff and delete, with source-linked recall back to their origin.

• ✓
  Pluggable storage backends

  SQLite for local single-writer use, PostgreSQL with pgvector for concurrency and scale, in-memory for testing.

• ✓
  Small auditable core

  The README puts the memory logic at about 500 lines, small enough for a security team to actually read.

• ✓
  Host adapters as sidecars

  Adapters run beside the desktop agent, slicing session logs into self-contained jobs without modifying the agent itself.

• ✓
  No LLM calls in the core service

  MemoryService handles storage, embedding and retrieval directly, so the memory layer adds no hidden model spend.

• •
  Carrying context between coding agents

  Move from Claude Code to Cursor mid-project without re-explaining conventions, architecture or past decisions.

• •
  Persisting project conventions as skills

  Repeated workflows become Markdown skill files the agent retrieves automatically on the next similar task.

• •
  Cutting repeated context tokens

  Retrieving a targeted memory entry costs far less than replaying an entire prior conversation into the prompt.

• •
  Auditing what an agent remembers

  Because memories are plain files, teams can review, redact or delete anything before it is reused.

• •
  Self-hosting memory for sensitive work

  Run the whole store on local SQLite or an internal Postgres so no session data leaves the network.

Self-hosting is free and the hosted memU Cloud is advertised as free with a login; there is no published paid tier or per-seat rate, and enterprise deployments are handled as scoped design-partner engagements, so budget engineering time rather than licence fees.

This page was written from 5 sources, 4 on domains other than memu.pro.

1. 1.github.com — memU
2. 2.raw.githubusercontent.com — README.md
3. 3.api.github.com — memU
4. 4.memu.pro — memu.provendor
5. 5.landscape.jimmysong.io — memu