AtomicMemory for Hermes Agent

AtomicMemory is a native Hermes memory provider. It participates in Hermes' memory lifecycle: background recall is prefetched for the next turn, completed turns are synced without blocking the chat loop, and the agent gets explicit tools for searching and storing durable facts.

By default, recall is shared across every AtomicMemory tool the user has touched (Claude Code, Codex, the web extension, etc.). Set memory_scope=siloed to restrict Hermes to Hermes-ingested memories only.

Architecture

Hermes Agent (Python)
  → $HERMES_HOME/plugins/atomicmemory/__init__.py
      → AtomicMemoryClient (Python protocol)
          → PythonSdkAtomicMemoryClient
              → published atomicmemory Python SDK MemoryClient

The Python provider owns Hermes lifecycle compatibility only — registration, hooks, tool schemas. Memory semantics flow through the published Python SDK.

Prerequisites

• Hermes Agent installed
• AtomicMemory core URL exported as ATOMICMEMORY_API_URL
• AtomicMemory bearer key exported as ATOMICMEMORY_API_KEY when using the Core Quickstart or any protected service

Install

Install the provider from the published npm package. The installer copies the Python provider files into Hermes' memory-provider directory; no repository clone is required.

npx -y @atomicmemory/hermes-plugin install
export ATOMICMEMORY_API_URL="http://127.0.0.1:17350"
export ATOMICMEMORY_API_KEY="local-dev-key"

Then select and verify the provider:

hermes memory setup
# select "atomicmemory"
hermes memory status
# confirm "atomicmemory" is active

For source development, symlink the checkout instead:

cd /path/to/atomicmemory
mkdir -p "$HERMES_HOME/plugins"
ln -s "$(pwd)/plugins/hermes" "$HERMES_HOME/plugins/atomicmemory"

Config

Hermes' setup wizard prompts for a minimal pair (scope_user, memory_scope). Advanced settings live in $HERMES_HOME/atomicmemory.json.

Connection details (ATOMICMEMORY_API_URL, ATOMICMEMORY_API_KEY) flow through environment variables into the Python SDK. The provider does not have a default API URL and fails to start if ATOMICMEMORY_API_URL is unset.

Environment

Provider-local file

$HERMES_HOME/atomicmemory.json accepts these keys (all optional):

Secrets are never persisted here — api_key and api_url are deliberately not in the allowed key set.

Memory scope

The source_site filter on recall is enforced through the Python SDK's AtomicMemory namespace handle. If the SDK is configured against a non-AtomicMemory provider (e.g. mem0), siloed mode fails loudly with PROVIDER_UNSUPPORTED rather than silently dropping the filter.

Memory mode

memory_mode selects which Hermes surfaces AtomicMemory exposes:

Tools

Lifecycle

• queue_prefetch(query) searches AtomicMemory in a background thread, with a generation counter so a slow earlier prefetch can't overwrite a faster newer one.
• prefetch(query) returns the most recent completed recall, then clears the slot.
• sync_turn(user, assistant) enqueues the turn to a single-writer worker thread and returns immediately. The worker calls client.ingest_messages(...) with provenance.source = "hermes" and provenance.sourceUrl = "hermes://session/<session_id>".
• on_session_end(messages) drains the worker, then closes the SDK client.

Reliability

A circuit breaker pauses SDK calls for two minutes after five consecutive failures and resets on the next success. Hermes continues to run while AtomicMemory is temporarily unavailable.

Troubleshooting

Tests

# Python provider and SDK adapter (deterministic, no network)
python3 -m unittest discover plugins/hermes/tests