In plain terms: AI coding assistants start every work session with a blank memory, so they can repeat a mistake that was already caught and fixed last month. My fix: keep a short “never do this again” list in a file the assistant is required to read before touching any code. Every mistake a human reviewer catches gets added to the list, so it can only happen once.

Agents forget between sessions; the repo doesn’t.

On my HRMS project, a seeder bug (duplicate employee IDs from an un-deduplicated pluck) silently broke CI — the automated pipeline that tests and ships every change. The fix was one line. The durable fix was a standing section in the repo’s CLAUDE.md that every agent session loads before touching code:

## Known CI Footguns — Never Re-introduce
1. Seeder dedup: always ->unique() on pluck('employeeId') in
   MockDataSeeder — duplicate IDs will break seeding silently
2. Leave calendar scoping: non-approver roles must only see
   their own leaves — global scoping breaks this
3. Staging Docker BuildKit cache is flaky — pre-existing,
   unrelated to code changes

The rule: every regression a reviewer catches becomes a constraint the next agent session starts with, instead of a mistake it gets to repeat.

Why this works better than a wiki:

  • The context file lives next to the code it protects, so it can’t drift out of the workflow — agents read it whether they “remember” to or not
  • Each entry is phrased as a constraint (“never re-introduce”), not a post-mortem — agents act on imperatives better than narratives
  • It stays short. If it grows past a screen, entries get promoted into lint rules or tests, which are even harder to ignore

What I learned

Don’t fight an AI assistant’s forgetfulness — design around it. The repo remembers so the agent doesn’t have to, and a mistake written down as a constraint (“never re-introduce X”) only ever happens once. The discipline that makes it work is keeping the list short: once an entry matters enough, it graduates into a test or a lint rule, which even a forgetful agent can’t walk past.

Related: hrma: an HR system built almost entirely by AI agents, Letting an AI hunt for the bugs humans skim past, Running several AI assistants at once