mem-search

Layered workflow. Do not dump full observations until IDs are filtered, and do not reach for raw tool bodies until the summaries came up short.

1. search(query) — index with IDs
2. timeline(anchor=ID) — nearby context
3. get_observations(ids=[...]) — full details for those IDs only
4. get_tool_uses(ids=[...]) — the ORIGINAL tool_input / tool_response for specific tool calls. Last resort: these are unsummarized and can run to thousands of tokens each.

The worker also live-writes a growing ID INDEX to agents/<this-seat-uuid>/memory/log/zz-claude-mem-inject.md (host Memory mid-attach). Use those IDs with get_observations. Thin seat diaries are filled from newest house observations. Disable with CLAUDE_MEM_GROK_BOT_INJECT_ENABLED=false.

Stamp platformSource as cursor or grok-bot on writes for this host. When reading, do not drop the other host unless asked.

If MCP is missing, run the install skill first (npx claude-mem install --ide <host>).