smriti

Smriti for Hermes

Smriti gives personal agents a local record of changing facts and routines. Ask what is true now or on a date, see the stored record behind the answer, and reuse the same memory from another local MCP host.

Validate the preview

The package targets Hermes Agent v0.23.0 with its MCP client enabled and requires python3 3.9 or newer with SQLite FTS5 on the host's PATH. From the assembled package directory, run:

hermes plugins validate "$PWD" --install-deps
python3 -E -B -S demo.py --portable

The 5,000-record replay seeds a synthetic history in a temporary directory, saves a weekday gym correction over MCP, restarts the server, and checks current and historical evidence. It reports warm local MCP timings and disk size. A live agent's model time and fees are separate.

If a minimal Hermes development runtime lacks the MCP client, prepare its official mcp extra using Hermes's package manager. A normal Hermes installation includes MCP support. Confirm that all six Smriti tools appear before running the conversation.

Try it in a conversation

Enable the Smriti plugin in a disposable Hermes profile. Ask Hermes to use Smriti to save a named routine, then correct it in the same scope. In a new session, ask for the current value and the value on a specified date. Inspect the returned source IDs and validity dates. Smriti exposes remember, recall, search, facts_about, add_fact, and stats over local stdio. The package also includes a personal memory workflow skill.

To film the dated example, run python3 -E -B -S demo.py --seed-db /absolute/path/to/demo-profile/plugin-data/resolved-directory/smriti.db before the MCP server opens that database. The parent directory must exist. This writes the synthetic history to a new file and refuses to overwrite an existing database. Then tell Hermes to save the weekday move to Oak Street Gym using the routine:weekday scope.

For another compatible local MCP client, use this directory as the server's working directory and the python3 launch arguments in mcp.json. Replace ${PLUGIN_DATA}/smriti.db with the absolute path to the same database if you want the clients to share memory. Use separate paths for separate stores. Share the database only with clients you trust to access its contents.

With evidence: true, recall returns compact source records and an omitted-record count. Fact records include scope, validity dates and temporal status. Episode records preserve their timestamp and speaker role. The char_budget bounds the serialized structured response. Plain recall keeps the core's existing context response.

Data and network use

Hermes starts the Python MCP subprocess directly. The launch fixes Smriti to offline lite mode, with no paid memory, embedding, or extraction service, no telemetry, and no credentials required. Memory stays in the active Hermes profile's plugin data directory as smriti.db and its SQLite sidecar files. The write tools persist the supplied facts or conversation entries. Supersession keeps previous facts for historical queries. An agent using a hosted model can send the tool results it receives to that model provider.

README from the reviewed commit 7735717; it updates when the author re-pins.