# Agent Cache Rules

This workspace is integrated with long-term semantic memory via `agentcache-python`.
You must act as your own memory manager by calling the cache MCP tools at critical boundaries.

## Rules & Workflow

1. **Initial Search (Prefetch Context)**:
   At the start of every session or new task, immediately call `cache_smart_search` with terms related to the current objective. This retrieves past architecture patterns, preferences, bug fixes, and past work.
   - Example: `cache_smart_search(query="jwt token rotation logic")`

2. **Log Observations**:
   As you work, call `agent_observe` to record what you are doing. Include `folderPath` (the working directory), `agentId` (your agent name), and `text` (what happened).
   - Example: `agent_observe(folderPath="/your/project/path", agentId="cursor", text="Refactored auth middleware to use JWT rotation")`

3. **Save Long-Term Memories**:
   When you successfully debug a complex error, discover an undocumented requirement, or establish a convention, persist it:
   - Call `cache_save` to save long-term structural facts. Always extract 2-5 specific lowercased tags (e.g. `auth-flow`, `refresh-token`) as concepts.

4. **Checklist Before Ending**:
   Before stating a task is complete:
   - Reflect on whether any insights should be saved via `cache_save`.
   - Review the folder activity feed with `cache_timeline` to confirm all significant work was captured.