Agent should decide when to save/recall memories.

[illeatmyhat]

See PR #266
In the hot path, agents should decide when memory is relevant because hooks interrupt the flow too much.
This means that for different agents, we need to find non-hook mechanisms to prompt the agent.

Bob

• Bob has rules, which are similar to AGENTS.md but can be loaded into individual files and are not tied to projects.
• Bob reads AGENTS.md but only for projects
• Bob has "modes" which essentially replaces the system prompt for agents.

Claude

• Claude has "output styles" which appends an excerpt into a system prompt. This is comparable to Bob's "modes"
• Claude reads AGENTS.md for projects and in ~/.claude/AGENTS.md, which is very convenient because that means it's always loaded.
• Claude has a built-in memory subsystem which cannot be turned off.

Codex

• Codex also reads ~/.codex/AGENTS.md but does not support @importing files for progressive disclosure.

Solution:
For bob, we use rules, and for Codex we create an EVOLVE.md file which AGENTS.md tells the agent to read prior to every conversation.
For Claude, we want to build on top of their memory system instead of trying to replace it or fight it. This may end up being a mistake as we have no ownership over it.
To that end, we create an adapter which transforms Claude memories into our preferred memory format so that we can stay consistent.