Great clarifying questions.
In practice, things usually break in this order:
- Commands and paths change. This is the most common case. Someone renames a script in
package.json, moves a config file, switches frompnpmtobun, butAGENTS.mdstill has the old instruction. Agents follow it with confidence, and you waste an extra cycle trying to figure out why the build command doesn’t work. AGENTS.mdgets outdated. This is second place. Users update code and forget the docs. The drift utility I mentioned earlier exists for exactly this reason, it’s the most boring and most common failure mode.- Old memories become wrong. It happens, but usually with a smaller blast radius. A memory like “we use Jest” survives a migration to Vitest and quietly poisons recommendations. It’s worth cleaning up memories every few weeks.
- Rules tied to specific tools drift apart. This is a slow burning issue. You tweak a rule for Cursor and forget to mirror it in the equivalent for Claude Code or Codex, and different tools start behaving differently. This is less “it broke” and more “why does Cursor do X, but Codex does Y”.
- Agents don’t load the right context. This is more of a constant background tax than a sudden break. Usually fixed with more explicit glob patterns in
.cursor/rules/*.mdcor by passing files directly in the prompt.
If I had to pick one rule of thumb, anything that duplicates info already in code or config will go stale over time. Anything that just points to that info tends to stay correct.
About MCP memory shared between tools, honestly right now it’s mostly experimental or one off setups. I see users run local stuff like Mnemosyne: A local, persistent memory MCP server for Cursor or Cursor-memory — Persistent, Searchable Memory for Cursor AI and get real value within a single workflow. Cross tool, reliable, team memory is a totally different beast. MCP support quality varies across tools, and memory that one tool writes and another reads only works if both are disciplined about when to write and when to retrieve, and in practice that isn’t there yet. I haven’t seen a setup I’d call production grade across multiple tools. If anyone reading has one, I’d genuinely love to hear about it.
For your workflow, the boring answer will probably win: AGENTS.md plus scripts as the source of truth, a small drift check in CI, clean up memories from time to time, and treat MCP memory as a nice bonus, not the foundation of the system.