Skip to content

Latest commit

 

History

History
65 lines (48 loc) · 2.84 KB

File metadata and controls

65 lines (48 loc) · 2.84 KB

Harness wiring notes

The kit is deliberately CLI-first: any harness that can run a shell command can use it. Specific notes per harness:

Claude Code

Two options, use both:

The hook can inject a relevant excerpt from another local harness. Registering mcp_server.py also gives Claude the explicit session_recall and session_list tools.

Hermes Agent

The reliable path is a skill. Create ~/.hermes/skills/<category>/agent-memory/SKILL.md with standard skill frontmatter (name, description stating when to use it) and the snippet from agent-instructions.md as the body, with the absolute paths to recall and remember on your machine. Hermes agents load skills before substantial work, so the description is what triggers usage - make it say "use at the start of substantial tasks to recall past fixes and findings".

mcp_server.py can also be registered if your Hermes version supports MCP servers - consult the Hermes documentation for the current registration syntax rather than guessing config keys.

Recent Hermes versions can run the same hook as a pre_llm_call shell hook. The hook detects Hermes's wire format and emits native context. Keep the explicit MCP tools as well so the agent can search deliberately.

Codex

Add the agent-instructions.md snippet to ~/.codex/AGENTS.md (or the project AGENTS.md) with absolute tool paths. Codex runs the CLIs via shell like anything else. Register mcp_server.py as a stdio MCP server when available so session recall is a first-class tool.

Grok

Register mcp_server.py as a stdio MCP server and set AGENT_MEMORY_CALLER=grok for clear usage receipts. Add the standing rule from agent-instructions.md to AGENTS.md, which Grok reads.

Do not install the Claude prompt hook as Grok auto-recall. Grok runs passive UserPromptSubmit hooks, but ignores their stdout, so the returned context never reaches the model. The supported continuity path is:

  • Grok transcripts are indexed from ~/.grok/sessions
  • session_recall is available through the agent-memory MCP server
  • the standing instruction tells Grok when to call it

Verify all three. A configured MCP entry alone is not proof.

Anything else with a shell

The contract is four commands: python3 ingest.py (rebuild documents), recall "query" -k 5 (search documents), session-memory recall "query" -k 5 (search dialogue), and remember --source X --evidence Y "text" (write a verified lesson). Wire them wherever your harness keeps standing instructions. For harnesses that speak MCP but not shell, use mcp.md.