Implement memory routing: global vs. project

[SandyTao520]

When the memory subagent saves something, it needs to put it in the right place:

• Global (~/.gemini/) - things about the user that apply everywhere: "I prefer concise commit messages", tool preferences, personal conventions.
• Project (.gemini/) - things specific to this codebase: "this project uses vitest not jest", architecture decisions, known pitfalls.

The subagent prompt needs clear routing guidance. There are genuinely ambiguous cases (e.g., "always use TypeScript strict mode" - is that a personal preference or a project convention?) so we need reasonable heuristics, not just hard rules.

The design doc specifically calls out needing "a small set of behavioral evals to lock in this behavior." We should have evals for: a clearly global fact, a clearly project-specific fact, and at least one ambiguous case to verify the heuristic.

[rnett]

I, and I suspect many others, have custom locations for memory files to support multiple agents and tools, storing in Git, etc. For a concrete example, I have used both a ~/.agents/AGENTS.md (that's picked up mostly automatically) and a ~/SomeProjectDir/my.agents/ROOT_AGENTS.md that is linked to ~/.gemini/GEMINI.md. It would be great if the memory agent would support these - I currently have a custom memory agent with the relevant info in its prompt and a skill that contains data on the whole setup.

The same applies to project configuration, although I suspect it's much simpler with the default usually being AGENTS.md.