All plugins

metronix

Metronix memory provider for Hermes with prefetch recall and built-in memory write-through.

mtrnix1MemoryUpdated 1 month ago

hermes-memory-metronix

Standalone Hermes MemoryProvider backed by Metronix Memory.

This repository owns the Hermes-native adapter layer:

  • Hermes plugin packaging
  • config and auth resolution
  • prefetch and write-through behavior
  • Hermes-focused unit, integration, and smoke tests

Metronix backend contracts and MCP integration docs remain in the
metronix-memory repository.

Status

This is a released standalone plugin. The current public package is
2026.34.1, and the
native Hermes smoke video
shows the end-to-end provider flow.

What it already does:

  • Implements the Hermes MemoryProvider contract
  • Prefetches relevant Metronix memory records before a turn
  • Mirrors Hermes memory(action="add") writes into Metronix
  • Optionally writes completed turns as session-scoped Metronix memory
  • Supports REST JWT or mtk_… personal-key auth, with email/password login fallback

Current scope:

  • No extra provider-specific model tools yet
  • Prefetch currently targets /api/v1/memory/search
  • Knowledge-document / page retrieval is left as a follow-up

Layout

The installable Hermes plugin is this repository root. Its provider
implementation lives in:

  • plugin/metronix/

Hermes discovers user-installed memory providers from:

  • ~/.hermes/plugins/<name>/__init__.py

Install

Requires Hermes and Python 3. Choose one option. Both use the same secure,
interactive setup command.

Option 1: install script

curl -fsSL https://raw.githubusercontent.com/mtrnix/hermes-memory-metronix/main/scripts/install-plugin.sh | bash

The script installs the latest main branch, then starts
hermes-metronix-setup.

Option 2: PyPI

uv tool install "hermes-memory-metronix>=2026.34.1"
hermes-metronix-setup --generate-token

--generate-token prompts for your Metronix email and password, then uses
curl to create a personal REST API key without printing it. If you already
have a REST JWT or personal API key, omit the flag and paste that token instead.

The setup command copies the provider to
~/.hermes/plugins/metronix, writes non-secret settings to
~/.hermes/metronix.json, stores the token in ~/.hermes/.env with 0600
permissions, and runs hermes memory setup metronix.

It never prints the token. Save the original token in a password manager and
rotate it if it is exposed. Do not use METRONIX_MCP_API_KEY: it is for MCP,
not the REST API used by this provider.

Release versioning

Public PyPI releases use YYYY.WW.REVISION: the calendar year and ISO week in
which the release is cut, followed by an incrementing revision. The latest
public release is 2026.34.1.

The matching GitHub release tag, for example v2026.34.1, points to the
release commit; that commit hash is the precise build identifier. Private
builds may append the hash as a PEP 440 local version, such as
2026.34.1+gabc1234, but local versions are not published to PyPI.

Hermes plugin manager

Alternatively, install this repository directly with Hermes:

hermes plugins install mtrnix/hermes-memory-metronix --no-enable

Quick smoke test

  1. Run hermes memory status; it should show Provider: metronix.

  2. Run hermes chat—there is no --memory-provider chat flag—and enter:

    Remember that my test marker is hermes-metronix-001.
    
  3. Start a fresh hermes chat session and ask for the test marker. A pass
    includes hermes-metronix-001 in the answer.

hermes memory status may display unused secret fields as “Missing”. The
provider accepts either METRONIX_AUTH_TOKEN, or both METRONIX_EMAIL and
METRONIX_PASSWORD.

Plugin config

The plugin reads non-secret config from:

  • $HERMES_HOME/metronix.json

And secrets from:

  • $HERMES_HOME/.env
  • process environment

Example metronix.json:

{
  "base_url": "http://localhost:8000",
  "workspace_id": "MTRNIX",
  "agent_id": "hermes",
  "prefetch": true,
  "prefetch_top_k": 8,
  "prefetch_types": ["fact", "preference", "pinned"],
  "cite_sources": true,
  "write_through": true,
  "write_scope": "workspace",
  "sync_turns": true
}

Secrets:

METRONIX_AUTH_TOKEN=<REST JWT or mtk_ personal key>
# or:
METRONIX_EMAIL=admin@metronix.local
METRONIX_PASSWORD=...

Important:

  • METRONIX_AUTH_TOKEN must be a REST JWT or mtk_… personal key for
    /api/v1/*. mtk_… keys require Metronix core's shared REST authentication
    resolver.
  • METRONIX_MCP_API_KEY is for /mcp, not /api/v1/memory/*
  • if you provide both a bearer token and login credentials, the client will
    retry once with a fresh login JWT when the original bearer gets a 401

Privacy and turn synchronization

sync_turns is enabled by default. When enabled, the provider sends the text
content of each completed user message and assistant response to the configured
Metronix endpoint and stores them as session-scoped memory records.

The provider does not currently transmit Hermes's complete messages
collection, tool-call arguments, or tool results. Prefetch search queries and
explicit Hermes memory writes are transmitted separately as part of the core
provider behavior described above.

To prevent completed conversation turns from being transmitted, set this in
$HERMES_HOME/metronix.json:

{
  "sync_turns": false
}

Mapping notes

Hermes write scopes map to Metronix scopes like this:

  • per_agent -> per_agent
  • workspace -> global
  • shared -> global
  • session -> session

The current prefetch path uses Metronix memory search, so prefetch_types
maps to Metronix memory kinds:

  • fact
  • preference
  • pinned

page is not wired yet because the provider does not call a unified
knowledge-search endpoint.

Migrating an existing llm-wiki into Metronix

Hermes's bundled llm-wiki skill maintains a directory of markdown pages
(entities/, concepts/, comparisons/, queries/, raw/) at $WIKI_PATH
(default ~/wiki). To bulk-ingest an existing wiki into Metronix's Knowledge
Base (chunked, embedded, searchable via metronix_search_fast/metronix_get):

python scripts/migrate_wiki.py \
  --wiki-path ~/wiki \
  --base-url http://localhost:8000 \
  --workspace-id MTRNIX \
  --auth-token "$METRONIX_AUTH_TOKEN"

Each page becomes a document with source_type="hermes_llm_wiki" and a
deterministic doc_label derived from its wiki-relative path, so re-running
the script updates existing documents instead of duplicating them. By
default, only raw/, entities/, concepts/, comparisons/, and
queries/ are ingested; SCHEMA.md, index.md, log*.md, and
_archive/** are skipped (pass --include-archive to include archived
pages instead).

Requires METRONIX_AUTH_TOKEN to be a REST JWT or mtk_… personal key for
/api/v1/*, not METRONIX_MCP_API_KEY, which is for /mcp only. mtk_…
keys require Metronix core's shared REST authentication resolver.

Development

Run the fast local checks:

uv run --extra dev pytest tests/unit -v
RUN_INTEGRATION_TESTS=1 uv run --extra dev pytest tests/integration/test_live_metronix.py -v

Verify against Hermes

Clone a clean Hermes checkout separately, then verify both the current
MemoryProvider abstract base class and Hermes's real plugin loader:

HERMES_AGENT_SRC=/absolute/path/to/hermes-agent \
  uv run --extra dev pytest \
  tests/unit/test_real_abc_contract.py tests/unit/test_hermes_plugin.py -v

The compatibility tests skip with an explicit reason when HERMES_AGENT_SRC
does not point to a Hermes checkout. They should pass before publishing a
plugin release.

README from the reviewed commit 55634c1; it updates when the author re-pins.