HarnessMemory for Hermes

Hermes MemoryProvider plugin backed by harness-memory.

For the full server-side OpenClaw + Hermes usage guide, see:

docs/harnessmemory-openclaw-hermes-usage.md

Install

From PyPI — pip install harness-memory-hermes pulls in harness-memory as a normal dependency, so this one command is enough:

pip install harness-memory-hermes

harness-memory-hermes install \
  --hermes-home ~/.hermes \
  --hermes-source ~/.hermes/hermes-agent \
  --force

harness-memory-hermes doctor \
  --hermes-home ~/.hermes \
  --hermes-source ~/.hermes/hermes-agent

From a monorepo source checkout, install this plugin package (editable) instead of PyPI:

python3 -m venv ~/.venv/harnessmemory
~/.venv/harnessmemory/bin/pip install -e ~/harness-memory/plugins/hermes

~/.venv/harnessmemory/bin/harness-memory-hermes install \
  --hermes-home ~/.hermes \
  --hermes-source ~/.hermes/hermes-agent \
  --source-dir ~/harness-memory/plugins/hermes/harnessmemory \
  --force

~/.venv/harnessmemory/bin/harness-memory-hermes doctor \
  --hermes-home ~/.hermes \
  --hermes-source ~/.hermes/hermes-agent

--source-dir only matters if you want to install from a specific checkout path rather than wherever pip put this package; it defaults to this package's own installed location.

The installer copies this provider to:

< HERMES_SOURCE >/plugins/memory/harnessmemory/

and activates it by setting:

memory:
  provider: harnessmemory

Manual install shape

If you do not use harness-memory-hermes install (e.g. Hermes' own native plugins install <git-url> flow), copy this directory into Hermes' memory plugin directory as:

< HERMES_SOURCE >/plugins/memory/harnessmemory/

Then activate:

hermes config set memory.provider harnessmemory

or edit ~/.hermes/config.yaml so memory.provider is harnessmemory.

Hermes' Python environment must also be able to import harness_memory. For source deployments, either install the project into Hermes' venv or add the source src/ directory through a .pth file.

Data path and config

By default the provider stores data under the active Hermes profile:

$HERMES_HOME/harnessmemory/memory.sqlite
$HERMES_HOME/harnessmemory.json

Do not hard-code ~/.hermes; Hermes profiles rely on hermes_home isolation.

Host Markdown indexing is separate from HarnessMemory's own SQLite store. By default the provider watches $HERMES_HOME/memories for Hermes built-in memory files:

• MEMORY.md
• USER.md
• allowlisted topical Markdown such as topics/*.md and projects/*.md

It does not scan $HERMES_HOME/state.db, sessions/, logs/, or arbitrary **/*.md.

harnessmemory.json supports the fields exposed through hermes memory setup:

{
  "profile": "balanced",
  "namespace": "hermes__default",
  "db_path": "$HERMES_HOME/harnessmemory/memory.sqlite",
  "recall": {
    "default_max_results": 5,
    "default_corpus": "all",
    "raw_policy": "fallback",
    "layer_order": ["atom", "page", "raw"]
  },
  "capture": {
    "include_roles": ["user", "assistant"],
    "min_message_chars": 0,
    "include_tool_calls": false,
    "include_tool_results": false,
    "skip_memory_echo": true,
    "host_files_watcher": true
  },
  "host_files_root": "$HERMES_HOME/memories",
  "host_files_allow": ["topics/*.md", "projects/*.md"],
  "privacy": {
    "redact_secrets": true,
    "redact_patterns": [],
    "store_raw_content": true,
    "store_tool_payloads": false
  }
}

capture and privacy apply to sync_turn(), on_memory_write(), and on_pre_compress() writes. recall applies to memory_search.

Exposed model tools

• memory_search
• memory_get

Provider CLI

When memory.provider=harnessmemory, Hermes exposes:

hermes harnessmemory status
hermes harnessmemory search "query" -n 5 --corpus all
hermes harnessmemory show raw/2026-06-04/evt_xxx.md --lines 40

These commands use the same memory_search / memory_get contract exposed to the model.

Lifecycle

• prefetch(query) recalls context before a model call.
• sync_turn(user, assistant) writes L0 raw events in a background thread.
• on_memory_write(action, target, content) mirrors built-in MEMORY.md / USER.md writes.
• on_pre_compress(messages) captures context before compression.