All plugins

dxmt

Long-term memory in a DXMT server you run: recall by meaning and keyword before each turn, durable facts extracted at session end and before compression, supersede and forget by id, and the same memories visible to every other agent and tool on the machine over REST and MCP. The server embeds, so no embedding key is needed in Hermes. Disclosure — conversation turns, extracted facts and built-in memory-tool writes are sent to the configured DXMT server (http://localhost:9999 by default; DXMT_URL or plugins.dxmt.server.url to change it) with DXMT_API_KEY if the server has one, read through Hermes; recalled memories come from that shared server, which other agents and tools on the machine can also write to; fact extraction sends session transcripts to Hermes's auxiliary LLM (task dxmt_extraction); `hermes memory setup` writes memory.provider and a plugins.dxmt block to config.yaml; no other endpoint, no telemetry, no shell access, no background process beyond prefetch and writer threads inside Hermes. Maintained by the DXMT project.

dxmtMemoryv0.1.0Updated 1 hour ago

About

Long-term memory in a DXMT server you run: recall by meaning and keyword before each turn, durable facts extracted at session end and before compression, supersede and forget by id, and the same memories visible to every other agent and tool on the machine over REST and MCP. The server embeds, so no embedding key is needed in Hermes. Disclosure — conversation turns, extracted facts and built-in memory-tool writes are sent to the configured DXMT server (http://localhost:9999 by default; DXMT_URL or plugins.dxmt.server.url to change it) with DXMT_API_KEY if the server has one, read through Hermes; recalled memories come from that shared server, which other agents and tools on the machine can also write to; fact extraction sends session transcripts to Hermes's auxiliary LLM (task dxmt_extraction); `hermes memory setup` writes memory.provider and a plugins.dxmt block to config.yaml; no other endpoint, no telemetry, no shell access, no background process beyond prefetch and writer threads inside Hermes. Maintained by the DXMT project.

hermes-dxmt-memory

DXMT, the open-source memory database for personal agents, as a
memory provider for Hermes Agent.

Hermes recalls from a DXMT server before each turn, by meaning and keyword, and extracts
durable facts at session end and before context compression. The agent gets four tools:
dxmt_recall, dxmt_remember, dxmt_read, dxmt_forget. The server does the embedding
(a local model through ollama or LM Studio, or any OpenAI-compatible endpoint), so Hermes needs
no embedding key, and the same memories are what every other agent on the machine sees over
REST and MCP.

What is specific to DXMT, against the memory providers that keep a store inside Hermes's
process or send turns to a hosted service:

  • One store, many agents. Memories are rows in a server you run. A fact Hermes keeps is
    recalled by anything else that talks to DXMT (OpenClaw, a LangChain app, curl), and a
    memory another tool wrote is recalled here when workspace isolation is off.
  • Supersede, not overwrite. dxmt_remember(supersedes=[id]) keeps the old fact as
    history and leaves it out of recall; the server deduplicates a fact said twice in the same
    category.
  • Keyword and meaning together. Recall is the server's hybrid search, so an exact name or
    error string is found as well as a paraphrase.
  • Nothing leaves the machine unless you point DXMT_URL somewhere else.

Installation: users

  1. Install Hermes Agent (see its README), then the plugin. The plugin lives in a subdirectory
    of the DXMT repository, which the installer understands:

    hermes plugins install dxmt/dxmt/vectordb/clients/hermes-dxmt-memory

  2. Install the client into Hermes's interpreter (shared across profiles, so once):

    uv pip install --python ~/.hermes/hermes-agent/venv/bin/python3 dxmt-python-sdk pyyaml

  3. Run a DXMT server with an embedding model. A local ollama pull bge-m3 is found on its
    own; DXMT_EMBEDDING_MODEL and DXMT_EMBEDDING_BASE_URL name any OpenAI-compatible one.

    docker run -d -p 127.0.0.1:9999:9999 -v dxmt:/data \
  --add-host=host.docker.internal:host-gateway -e OLLAMA_HOST=host.docker.internal:11434 dxmtai/dxmt

  4. Activate it and pick dxmt:

    hermes memory setup
hermes memory status      # Provider: dxmt

The server is elsewhere? DXMT_URL=http://host:port in the environment, or
plugins.dxmt.server.url in ~/.hermes/config.yaml. An API key, if the server has one, is
read from DXMT_API_KEY. Try it in an isolated profile first with hermes profile create dxmt-demo and -p dxmt-demo on every command.

What gets stored

| Hermes event | DXMT memory |
| --- | --- |
| dxmt_remember, extraction, built-in memory tool mirror | kind = category (preference, entity, event, case, pattern, general); abstract, provenance, origin, session_id in metadata |
| each completed turn (turns.enabled) | kind turn, id derived from (session, index, role, text) so a replay writes nothing twice |

Every row carries the tags hermes, hermes:ws:<workspace>, hermes:user:<id>,
hermes:agent:<identity>. Recall filters on the workspace tag while
retrieval.isolate_workspace is true (the default); set it false to recall across workspaces
and across the other tools writing to the same server.

Configuration

Defaults are in src/default_config.yaml; override under
plugins.dxmt in ~/.hermes/config.yaml:

plugins:
  dxmt:
    server:
      url: http://localhost:9999   # DXMT_URL wins over this
      api_key_env: DXMT_API_KEY
    retrieval:
      top_k: 10                    # dxmt_recall default
      prefetch_k: 5                # memories injected before a turn
      min_score: 0.0               # cosine similarity floor (1.0 = identical)
      isolate_workspace: true
    turns:
      enabled: true
    extraction:
      enabled: true                # LLM extraction at pre-compress and session end
      min_turns: 3

Extraction uses Hermes's auxiliary LLM (call_llm, task dxmt_extraction), so it follows
the model and limits configured there.

Development

cd vectordb/clients/hermes-dxmt-memory
ln -sf "$PWD" ~/.hermes/plugins/dxmt          # Hermes loads the directory; edits apply next session
HERMES_AGENT_DIR=/path/to/hermes-agent pytest  # offline: the server is a stand-in
DXMT_TEST_URL=http://127.0.0.1:9999 HERMES_AGENT_DIR=... pytest -m live   # against a server with a model

The offline suite covers the full tool loop, scoping, idempotent turns, extraction and
provenance against an in-memory stand-in for the server's memory API; the live suite drives the
same loop through Hermes's own MemoryManager.

Disclosure

Conversation turns, extracted facts and the built-in memory tool's writes are sent to the
configured DXMT server, http://localhost:9999 by default. Nothing is sent anywhere else.

README from the reviewed commit aaf7aa6; it updates when the author re-pins.