Repository navigation
The maintainer's side: several machines, several agents, one memory #1428
doobidoo
started this conversation in
Show and tell
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
It started on 26 December 2024, a month after Anthropic published MCP. Claude had no memory at all back then, and the chat apps that had one kept it locked inside their own product. What I wanted was a tool for everyday project work: somewhere to put down all the different threads and pick them up again when I need them. v0.1.0 went out the next day, ChromaDB and sentence-transformers behind Claude Desktop. Context injection followed in April 2025 with claude-memory-context, which wrote what the service remembered into Claude Desktop's project instructions, so a new chat started with the context already in place. In August 2025 mcp-context-provider took over the static part: tool rules and conventions that every session should get, while the memory service keeps what changes from day to day.
The database I use today goes back to June 2025. It holds 8,593 memories, 331 of them from this month, and I typed very few of them myself. They come out of Claude Code, OpenCode, Codex or Hermes sessions working on this repo and other stuff: a decision made in a review, a root cause found late at night, a CI trap that cost an afternoon. About 730 are typed as decisions, learnings, bugs or gotchas. Those are the ones that earn their keep.
The setup is the hybrid backend. SQLite-vec locally on each machine, Cloudflare D1 and Vectorize syncing in the background. A session-start hook pulls the relevant memories into every new Claude Code or OpenCode session before I write the first prompt, and the repo's CLAUDE.md gives every agent the same lookup order: files first, memory second, me last.
That runs on several PCs and notebooks, at home and at work. Because reads and writes hit the local database first, it also works offline. On a plane the agents keep remembering, and once there is network again the sync catches up.
For me the change is consistency more than speed. A triage decision from three weeks ago is still there when the topic comes back, so the agent does not contradict what I told a contributor last month. When I paste a message from Claudio, the session already knows who he is, which parts of the code he owns, and what we agreed on. I often have three sessions open on the same checkout. They don't share a context window. They do share what has been decided.
It has also bitten me. In February a test cleanup deleted 8,663 production memories in a single run. An emergency backup brought them back, and the test suite now refuses to start when it sees a production path. Nothing shows you how much you depend on a tool like watching it go empty.
The limit is the obvious one: memory is only as good as what goes in. 60 percent of my entries are plain observations, and plenty of those are noise that quality scoring and consolidation still have to sort out. Search finds the right decision most of the time. Not every time.
Claudio's post in #1427 covers the multilingual, many-agent side. This is the maintainer's side: one person, several machines, several agents, and one memory they all read from.
All reactions