The memory subagent (from #22716) needs evals to make sure it actually does the right thing and doesn't regress as we iterate. The design doc calls out several specific behaviors we need to lock in:
- Adding memories – agent saves relevant facts when the user states a preference or the session produces something worth remembering.
- Removing stale memories – when the agent touches a memory file, it notices and cleans up entries that are no longer true.
- De-duping – agent doesn't create a second entry for something it already knows. It consolidates instead.
- Organizing + TOC updates – GEMINI.md stays clean and structured as a table of contents; the agent updates it when memories change.
- Routing – user preferences go to
~/.gemini/, project knowledge goes to .gemini/. The doc specifically flags this as needing evals.
These should live in evals/ alongside existing behavioral evals. At least one eval per category.
The memory subagent (from #22716) needs evals to make sure it actually does the right thing and doesn't regress as we iterate. The design doc calls out several specific behaviors we need to lock in:
~/.gemini/, project knowledge goes to.gemini/. The doc specifically flags this as needing evals.These should live in
evals/alongside existing behavioral evals. At least one eval per category.