The author developed an open-source project called MEX that uses Git as a shared memory layer for coding agents, allowing context to be version-controlled and shared among engineering teams.
Hello! I’ve been working on shared memory for coding agents lately. its an open source project, link for the repo is in the replies The obvious way to build it is some shared backend that every agent talks to. Database, sync service, accounts, permissions, all of that. But the more I thought about what team memory actually needs, the more it started sounding like stuff Git already does. You want history. You want diffs. You want changes to move with the repo. You want branches to carry their own state. You want conflicts to be visible instead of silently overwriting each other. So I tried building the shared memory layer around Git instead. The basic setup is pretty simple. Canonical project memory lives as normal files in the repository. Architecture, decisions, specs, workstreams, handoffs, activity, etc. get committed like code. Each checkout keeps its own local indexes for search and code intelligence. Those don’t get shared. If another engineer pulls the repo, they rebuild the indexes against their own checkout. So the rough model is: The part I find most interesting is that the memory now has the same history as the code it describes. If someone changes a project decision, you can review the diff. If two people change the same piece of shared context, Git exposes the conflict. If an agent learns something useful, that doesn’t have to disappear inside one chat session. I also started using the same idea for handoffs. Instead of ending a session with a giant chat summary, an agent can prepare a structured handoff with completed work, blockers, decisions, changed files, next actions, and the repo state it was written against. The sender commits it, the next person pulls it, and it becomes part of the project history. One thing I did not want was agents silently turning whatever they wrote into shared truth. So some shared changes use a proposal flow. The agent can prepare a local draft, but publishing and accepting a Spec are separate actions that require explicit approval. I’ve been calling this Git-native team memory. It’s part of the open-source project I’ve been building, MEX Still early, and there are obvious tradeoffs. Git is not a realtime message bus, MEX does not auto push or pull anything, and two disconnected clones can still create normal Git conflicts. But I’m starting to think project memory should behave a lot more like source code than like another SaaS database. Would genuinely love to hear how other people are thinking about shared memory between coding agents, especially if you’ve tried solving this with Git, a database, MCP, or something else.
The author open-sourced o8, an MIT-licensed orchestrator that manages multiple coding agents in isolated git worktrees, with a merge gate, audit trail, and mobile approvals.
The author built mex, a tool that helps coding agents preserve repo knowledge instead of re-reading, with v0.7.0 adding a local code graph that reduces context usage by 90.7% while maintaining recall.
The author argues that spreadsheet agents are slow to adopt because they lack Git-style collaboration infrastructure (diffs, reviews, rollbacks) that makes coding agents usable. They announce an early runtime to address this gap.
The author built Nice Coding Agent, an open-source coding workbench with a visible and editable context stack, allowing users to curate exactly what the LLM sees. It features local-first retrieval, sandboxed execution, and hybrid code search, aiming to give developers control and visibility over context assembly.
I work solo across a lot of repos, and I keep hitting the same wall with Claude Code. Two separate problems that compound: Context dies on compaction. Long session, lots of hard-won detail about why something is built the way it is, then it compacts and most of that is gone. The only thing that reliably survives is whatever I wrote to a file mid-session. Memory is scoped per project folder. Each repo gets its own isolated memory keyed to its path. So conventions I've established, infrastructure