Here is the small-team problem with AI coding assistants: you spend a session getting your agent to actually know your project, then a colleague opens the same codebase with a different tool, or you start a fresh session the next morning, and the agent is back to knowing nothing. Every team has this. Most teams have just accepted it.
What's happening
akitaonrails/ai-memory shipped v3.1.1 two days ago, and the headline feature is one that small teams with mixed tooling will feel immediately.
The project stores AI coding agent memory as a git-backed markdown wiki inside your repository. When a session ends, lifecycle hooks compile what was discussed, decided, and left pending into wiki pages. When the next session starts, the agent reads a brief before touching any code. It is the same idea as having a project notebook that everyone on the team writes into after each work session, except the agent does the writing automatically.
Version 3.1 added cross-runtime portability. The wiki now works across more than 20 agent runtimes: Claude Code, Cursor, Codex CLI, Hermes Agent, and most others in active use. The project memory is not tied to one tool. A developer using Cursor and a contractor using Claude Code read from and write to the same wiki, because it lives in the repository, not in any vendor's cloud.
The memory is plain markdown. Version-controlled. Searchable with grep. No database, no API key, no account. The project has roughly 10,000 GitHub stars and has been in active development for about eight months.
Try this this week
- Visit github.com/akitaonrails/ai-memory and read the README. Setup is a few lines added to your agent config file and a shell alias for the lifecycle hooks.
- Start a coding session as usual. The hooks run silently in the background.
- At the end of the session, run the close hook. Check the wiki directory for the new entry.
- Open a fresh session tomorrow. Ask your agent to read the project brief before you start work. See whether it already knows what you were working on.
- If a colleague works on the same repo, have them run the same setup. They inherit all the context already in the wiki, regardless of which agent they use.
The bigger picture
The most durable AI setup for a small team is one where you own the data. Project memory stored as plain markdown inside your own repository means no dependency on any one tool's cloud, no subscription that disappears when a startup pivots, and no vendor migration when you switch tools next year. For a small shop running lean, that is not a technical preference. It is a survival instinct. The smallest working thing, and the one you actually keep.