Skip to content

total-recall in Gemini CLI

Status

  • MCP support: yes (stdio transport, via ~/.gemini/settings.json or gemini mcp add)
  • Hook support: no
  • Session storage: ~/.gemini/tmp/<sessionId>/logs.json (replay-style)
  • Adapter complexity: ~580 LOC (replay reconstruction + tool-call translation)

Install the MCP server

  1. Install total-recall so python -m total_recall.mcp_server works:
pip install total-recall
  1. Either drop this into ~/.gemini/settings.json:
{"mcpServers":{"total-recall":{"command":"python","args":["-m","total_recall.mcp_server"],"timeout":30000,"trust":false}}}

…or use the CLI helper:

gemini mcp add -s user total-recall python -m total_recall.mcp_server
  1. Restart Gemini CLI (the MCP registry is read at startup).
  2. Verify: gemini mcp list should show total-recall.

What you get

  • 26 MCP tools (recall, get_operator_context, check_banned, ...).
  • No hook integration — Gemini CLI has no hook surface yet.

Session ingest

total-recall autodetects Gemini CLI sessions by walking ~/.gemini/tmp/<sessionId>/logs.json and replaying the event log into canonical records. Verify:

total-recall sources test gemini_cli

Caveats

  • trust: false keeps the server sandboxed — recommended. Set true only if you understand the security implications.
  • The replay layer reconstructs message/part order from event timestamps; sessions with clock skew may end up with slightly out-of-order assistant turns.
  • Function-call and function-response events are translated into tool_use / tool_result blocks so downstream extractors don't need a special case for Gemini.