mex makes AI multiplayer. What one engineer's agent learns, every agent on the team can use. The memory lives in your repo, right next to the code.
25,000+developers trust mex globally
Open-source CLI: npx mex-agent@0.8.3 setup Try it on GitHub · Join the Discord
Install the CLI · What project memory is · OSS vs hosted · Hosted mex · CLAUDE.md vs memory
On a team, each engineer's agent learns the codebase alone. Memory fragments across machines, drifts from the real code, and nobody can see what the agents think is true.
01 / Fragments
Each developer's agent learns the codebase on its own, so no two share the same memory. Two agents solve the same problem different ways, and the conflict only surfaces in review, after it's already in the PR.
02 / Goes stale
The codebase changes every day; the per-developer memory does not. Agents keep enforcing rules that no longer match what's shipped, and nothing flags the gap.
03 / Ungoverned
There's no shared source of truth and no way for a lead to review, correct, or govern the memory every agent is working from. The team's standards end up living in tools leadership can't see into.
Canonical memory lives in the repository as readable Markdown and travels with Git. Each teammate keeps local indexes and a local Hub on their machine. Agents read and help maintain that memory through project instructions and the CLI. Works with Claude Code, Codex, Cursor, and other tools the team already uses.
For teams that want one shared deployment and an admin surface to govern what gets promoted, request early access. That product is not a rename of the CLI.
Admin dashboard
governs sync, drift, access, and policy
Hosted early access
Hosted early access (waitlist)
See drift, review what gets promoted into shared memory, and keep team context honest. This surface is the waitlist product.
mex is not locked to one coding agent. Memory lives with the project.
mex is open source and moving fast. Here's what's new in the CLI, and how we think about memory for AI coding teams.
Graph refresh reuses unchanged files, grounding survives fresh clones and small renames, and FastAPI, Flask and NestJS routes join the graph. npx mex-agent@0.8.3 setup
Setup starts in the local browser Hub. Terminal setup remains available.
Context graph in Hub, Relays usable before you know who picks up the work, compiler work no longer blocks Hub during graph construction.
After 367 commits of real code changes, memory tools with no link to code served 90 to 100% of outdated facts as current in their no-LLM modes. mex served 20%, and we measured what its warnings cost.
Read moreThere is a kind of sentence that sounds like knowledge but does not point at anything, and a name can survive the thing it named.
Read on LinkedInThoughts on agent memory. Clark and Chalmers’ Extended Mind Thesis holds up for this space as well.
Read on LinkedInOn a set of retrieval tasks, mex returned answers that were 10x more compact than a grep-based baseline, while still finding everything it needed to find.
Read moreEvery few months, a new model arrives with a larger context window. But being present is not the same as being used.
Read on LinkedInWhy every session starts cold, and the case for persistent, structured memory over giant context files.
Read on XWe started like most teams: every developer keeping their own rules file. Agents drifted apart within weeks. Corrections piled up, context went stale, and nobody had a clear view of what the tools actually knew.
So we built mex: project memory that lives with the code. The open-source CLI is that memory in your repo, today. The hosted product on this waitlist is for teams that want it governed in one place.
From single-player agents to multiplayer AI: one shared source of truth.
