Graphnosis memory provider for Hermes Agent

Local, encrypted, on-device memory for Hermes Agent via the Graphnosis desktop app — engrams, semantic recall, auto-prefetch, and (optionally) skills. No API key, nothing leaves your machine.

This is a standalone plugin, published separately per the Hermes memory-provider policy (plugins/memory/ is closed to new in-tree backends).

Prerequisites

1. Install Graphnosis and unlock your cortex (the app must be running).
2. Confirm the MCP socket exists: ~/.graphnosis/mcp.sock.
3. A working Hermes Agent install.

Install

Option A — pip + installer (recommended)

pip install hermes-graphnosis
hermes-graphnosis install          # copies the plugin into ~/.hermes/plugins/graphnosis/

hermes-graphnosis also supports status, uninstall, path, and install --force. It honors $HERMES_HOME (defaults to ~/.hermes).

Why an installer, not a pip entry point? Hermes (as of 0.17.0) discovers memory providers only by scanning plugin directories — see plugins/memory/__init__.py. A pip entry point for a memory provider is detected by the general PluginManager but coerced to kind=exclusive and handed to memory discovery, which never enumerates entry points — so it is silently dropped. The installer copies the plugin into the directory Hermes actually scans.

Option B — drop-in directory (no pip)

git clone https://github.com/nehloo-interactive/hermes-graphnosis
cp -R hermes-graphnosis/src/hermes_graphnosis/graphnosis ~/.hermes/plugins/graphnosis

Activate

hermes memory setup            # select "graphnosis"
hermes graphnosis status       # verify socket + cortex connectivity

Start a new Hermes session after setup.

Recommended: also install the MCP catalog entry

The memory provider gives you graphnosis_recall, graphnosis_remember, and graphnosis_stats. For the full tool surface — edit, forget, cross_search, skills, consent — add the Graphnosis MCP server. The manifest is bundled at src/hermes_graphnosis/data/mcp-manifest.yaml; register it in ~/.hermes/config.yaml:

mcp_servers:
  graphnosis:
    command: npx
    args: ["-y", "@graphnosis/mcp-relay", "${HOME}/.graphnosis/mcp.sock"]
    enabled: true

These appear as mcp_graphnosis_* tools in Hermes sessions.

Optional recall-hygiene skill:

hermes skills install https://graphnosis.com/skills/graphnosis/SKILL.md

Configuration

Settings live in $HERMES_HOME/graphnosis.json:

Key	Default	Description
socket_path	~/.graphnosis/mcp.sock	Graphnosis MCP socket path
default_engram	""	Default engram for graphnosis_remember
prefetch_max_tokens	1500	Prefetch token budget per turn

CLI

hermes graphnosis status
hermes graphnosis test-recall "project priorities"

Available only when memory.provider: graphnosis is active.

Troubleshooting

Symptom	Fix
Socket not found	Open Graphnosis and unlock the cortex; check ~/.graphnosis/mcp.sock
Recall returns "locked"	Unlock the cortex in the app
Provider inactive	Set memory.provider: graphnosis in ~/.hermes/config.yaml
hermes graphnosis unknown	Re-run hermes-graphnosis install; start a fresh session

Full docs: graphnosis.com/getting-started/connect-ai

License

Apache-2.0. See LICENSE.