rushdb-hermes-memory

Native RushDB persistent memory for Hermes Agent.

The provider recalls scope-authorized historical context before inference, captures completed turns without blocking the response, and uses a profile-local durable outbox so acknowledged events survive process restarts.

Requirements

• Python 3.11–3.13
• Hermes Agent with memory-provider entry-point support
• A RushDB project API key from app.rushdb.com, or a standalone RushDB engine and an embedding provider (see below)

Install

Install the provider into the same Python environment as Hermes Agent:

pip install rushdb-hermes-memory
hermes memory setup

Set the RushDB project key in the Hermes environment:

export RUSHDB_API_KEY="your-project-api-key"

Optional environment variables:

• RUSHDB_API_URL selects a different RushDB Cloud API endpoint. To self-host, run the RushDB engine and use the settings below.
• RUSHDB_MEMORY_SCOPE_SALT makes participant hashes deployment-specific.

A standalone RushDB engine

Recall is a vector search. RushDB Cloud embeds memories for you; a standalone engine (rushdb) does not, so the provider embeds them itself through an OpenAI-compatible embeddings endpoint:

export RUSHDB_ENGINE_URL="http://127.0.0.1:3000"
export RUSHDB_ENGINE_TOKEN="your-engine-token"
# export RUSHDB_ENGINE_SCOPE="tenant-1"   # only for an engine started with --multitenant
export RUSHDB_EMBEDDING_API_KEY="sk-..."
export RUSHDB_EMBEDDING_MODEL="text-embedding-3-small"
export RUSHDB_EMBEDDING_DIMENSIONS="1536"
# export RUSHDB_EMBEDDING_BASE_URL="https://api.openai.com/v1"

On start the provider creates the external vector indexes it needs; every memory is written with its vector and recall searches with the embedded query. With RushDB Cloud the embedding settings are optional. When the configuration is incomplete, hermes memory status reports what is missing.

Hermes discovers the rushdb provider through the hermes_agent.memory_providers Python entry-point group.

Behavior

• Recall searches EPISODE.summary and active MEMORY_FACT.text records with mandatory agent, profile, privacy, participant, and sandbox scope filters.
• Recall has a short fail-open timeout and falls back to bounded recent in-process episodes.
• Successful primary-agent turns are captured asynchronously as canonical EPISODE records.
• Explicit Hermes memory add and replace operations are mirrored as active MEMORY_FACT records.
• Writes first enter an atomic local outbox under $HERMES_HOME/rushdb-memory/outbox, then use deterministic IDs and idempotent RushDB upserts.
• Surviving outbox files are retried while the process is running and replayed after restart.

Privacy and trust

The provider uploads bounded user and assistant text from completed primary-agent turns. It does not automatically upload Hermes' complete messages payload, system prompts, tool calls, command output, or local file paths. Subagents do not write memory.

Recalled records are formatted as untrusted historical context, never as instructions or policy. Participant identifiers are stored as hashes rather than raw platform IDs.

Development

uv sync --locked --group dev
uv run ruff format --check .
uv run ruff check .
uv run pytest
uv build

Release

CI verifies pull requests and every push to main. A successful main build publishes the version in pyproject.toml to PyPI when that exact version is not already present. Configure the pypi GitHub environment and its PYPI_TOKEN secret. Bump version before merging a subsequent release.

License

Apache-2.0