agentmem

Agent Memory Instructions

Copy-paste the block below into your CLAUDE.md, .cursorrules, or AGENTS.md file. This tells your coding agent how to use agentmem effectively.


For Claude Code (CLAUDE.md)

Paste this into your project’s CLAUDE.md or ~/.claude/CLAUDE.md:

## Memory (agentmem)

You have access to a governed memory system via MCP tools (prefixed `mcp__agentmem__`).

### Session Protocol
- **Start of every session:** call `load_session` to restore context from the previous session.
- **End of every session:** call `save_session` with: what was done, what's in progress, what's blocked.

### Before Acting on a Remembered Rule
- Call `search_memory` to verify the rule is still current (status = active or validated).
- Do not act on deprecated or superseded memories.

### When You Learn Something Durable
- Call `search_memory` first to check for duplicates.
- If no match, call `add_memory` with the appropriate type (bug, decision, setting, procedure, context, feedback).
- New observations start as `hypothesis`. Promote to `active` once confirmed.

### Trust Hierarchy
- `validated` > `active` > `hypothesis`. Deprecated and superseded are excluded from recall.
- Only the user promotes to `validated`. You may add `hypothesis` or `active`.

### Periodic Health
- Run `memory_health` when starting a new task area. Flag any conflicts or stale memories to the user.
- Run `memory_conflicts` if you notice contradictory rules.

For Cursor (.cursorrules)

Paste this into your project’s .cursorrules file:

## Memory (agentmem)

You have governed memory via MCP tools.

Session protocol:
- Start: call load_session to restore prior context.
- End: call save_session with status of current work.

Before acting on a remembered rule:
- Call search_memory to verify it's still active or validated.
- Ignore deprecated or superseded memories.

When you learn something durable:
- Search first (no duplicates).
- Add with type: bug, decision, setting, procedure, context, or feedback.
- New observations = hypothesis. Promote to active once confirmed.

Trust: validated > active > hypothesis. Deprecated excluded from recall.
Run memory_health at session start. Flag conflicts to the user.

For Codex CLI (AGENTS.md)

Paste this into your project’s AGENTS.md:

## Memory (agentmem)

Governed memory is available via MCP tools.

- **Session start:** `load_session` to restore context.
- **Session end:** `save_session` with what was done, in progress, and blocked.
- **Before acting on memory:** `search_memory` to verify the rule is active/validated.
- **When learning something new:** `search_memory` first (no duplicates), then `add_memory`.
- **Trust:** validated > active > hypothesis. Deprecated/superseded excluded.
- **Health:** Run `memory_health` periodically. Flag conflicts to user.