Git as an Agent Memory: Versioned Lessons a Team Can Audit
Learning you cannot inspect is not learning. It is drift. Git turns an agent lesson into a record.
People want AI agents that get better at their recurring work. The dangerous version of that is an agent that silently rewrites its own instructions after every job. One bad answer, and every future job inherits the mistake with no way to see where it came from. The safe version keeps a knowledge vault in Git: lessons as reviewed, versioned files.
What a lesson looks like
A lesson is not a vibe. It is a small record with a title, the problem it addresses, the method that worked, the signals that tell you it applies, the hard "no" list, and the evidence it was proven. Stored as a file under a dated folder, it can be read, diffed, and rolled back like any other code.
A worker can discover a better method. It cannot install one into its own brain unreviewed.
The review gate
When a worker finds a reusable improvement, it files a candidate revision — old instruction, proposed instruction, reason, evidence, affected templates — into a queue. It does not become active until a human or the master model approves it. Approved lessons are exported to the Git-tracked vault; rejected ones stay in history as a record of what did not hold up. Nothing self-promotes.
Why Git specifically
Git already solves the hard parts: versioning, diffs, blame, rollback, and a shared source of truth a team can review. Putting agent lessons there means the improvement is recoverable if a machine dies, auditable if someone asks "why did the agent start doing this," and safe to revert if it turns out wrong. Private job history, credentials, and raw logs stay out of Git — only the redacted, approved lesson goes in.
The build-to-lesson loop
The pattern also runs the other direction. When the master model lands a strong result, the reusable part is distilled into a lesson the smaller workers can follow. The master does the hard version once; the workers inherit the checklist. That is controlled learning — not model retraining, but a versioned playbook that makes cheap agents reliable at your specific tasks.
The takeaway
If your agents "learn," ask where the lesson lives and who approved it. If the answer is "somewhere in the model, nobody knows," that is drift. If the answer is "a reviewed file in Git," that is a record you can trust.