All plugins

elm

Local, reviewable long-term memory. You save notes (or approve notes the model proposes) and matching notes are recalled automatically. Off until the owner turns it on from a terminal: for all Desktop/TUI/CLI chats at once (hermes elm auto on, notes shared across chats; any chat can opt out with /elm private) or per chat (hermes elm grant). The model can never turn it on. Works as a general plugin or in the memory slot (memory.provider: elm, with the recall indicator; hermes memory setup elm offers auto mode). Optional memory-aware delegation backend (delegation.backend: elm): you name agents (models or profiles) and ELM routes subagent tasks to the one that did best on similar earlier tasks. Disclosure — no network calls, dependencies, telemetry, self-updater or background daemon of its own. Data stays in the profile's plugin-data/elm SQLite file. Recalled notes are added to the current message and so reach your model provider; in auto mode notes saved in one chat are recalled in your other chats (never in messaging, cron, subagent or ACP chats). Two optional, per-chat opt-in features make model calls through Hermes's auxiliary model slots: a search planner (query text only) and an organizer (selected captured text). Opt-in capture stores finished turns as plain text locally. Routing stores agent names, up to 12 plain-word goal keywords, success and duration (no task or result text), only for chats with ELM on, scoped like notes. Deleting a chat deletes its chat-only notes, captures and routing history (shared notes stay). hermes elm sessions reads the profile's session list to help pick a chat ID.

Miawd3Memoryv0.4.0-previewUpdated 3 days ago

About

Local, reviewable long-term memory. You save notes (or approve notes the model proposes) and matching notes are recalled automatically. Off until the owner turns it on from a terminal: for all Desktop/TUI/CLI chats at once (hermes elm auto on, notes shared across chats; any chat can opt out with /elm private) or per chat (hermes elm grant). The model can never turn it on. Works as a general plugin or in the memory slot (memory.provider: elm, with the recall indicator; hermes memory setup elm offers auto mode). Optional memory-aware delegation backend (delegation.backend: elm): you name agents (models or profiles) and ELM routes subagent tasks to the one that did best on similar earlier tasks. Disclosure — no network calls, dependencies, telemetry, self-updater or background daemon of its own. Data stays in the profile's plugin-data/elm SQLite file. Recalled notes are added to the current message and so reach your model provider; in auto mode notes saved in one chat are recalled in your other chats (never in messaging, cron, subagent or ACP chats). Two optional, per-chat opt-in features make model calls through Hermes's auxiliary model slots: a search planner (query text only) and an organizer (selected captured text). Opt-in capture stores finished turns as plain text locally. Routing stores agent names, up to 12 plain-word goal keywords, success and duration (no task or result text), only for chats with ELM on, scoped like notes. Deleting a chat deletes its chat-only notes, captures and routing history (shared notes stay). hermes elm sessions reads the profile's session list to help pick a chat ID.

ELM — long-term memory for Hermes (preview)

ELM gives Hermes a local, reviewable long-term memory. It is a normal Hermes plugin: no changes to Hermes itself.

  • You decide what is remembered. Save notes yourself, or let the model propose notes that you approve.
  • On for all your chats, if you want. One terminal command (hermes elm auto on) and every chat remembers what you saved in any other chat. Any chat can opt out with /elm private.
  • Recalled automatically. Relevant approved notes are added to the model's context when they match what you ask.
  • Local. Everything is stored on your computer, in your Hermes profile. Nothing is uploaded except what ends up in a chat you send to your model provider.

Preview. This build is for testing. Please use the tester guide and report what you find.

Requirements

  • Hermes 0.23.1 or newer (Desktop, TUI or CLI), on Windows, macOS or Linux.
  • A model configured in Hermes. ELM itself works without any model; the optional search planner and organizer use one.

Install

hermes plugins install "https://github.com/Miawd3/ELM-Hermes-plugin.git#stock-plugin"
hermes plugins enable elm

Or in Desktop: Plugins → Install from repository, paste the URL above.

Restart Hermes. In Desktop, open Capabilities → Plugins and make sure ELM is switched On.

Check it: hermes elm doctor should print Storage: ok.

Optional: make ELM Hermes's memory provider

Hermes has one slot for an external long-term memory. To put ELM there:

hermes memory setup elm        # or: hermes config set memory.provider elm

Setup also offers to turn ELM on for all your chats (auto mode, below). Restart Hermes. hermes elm doctor then says Memory slot: ELM is the memory provider, and chats show
🧠 ELM — recalled N memories when a note is added. Grants, /elm commands and privacy rules are
the same in both modes, and a note is never added twice. Keep the plugin enabled
(hermes plugins enable elm): /elm, hermes elm grant and routing come from the plugin side.
Hermes's own built-in memory keeps working next to it. To go back:
hermes config set memory.provider builtin, then restart Hermes.
Only one external memory provider can be active, so this replaces another one (for example Honcho or mem0) if you had it.

Turn ELM on

ELM never turns itself on, and the AI can never turn it on. You choose, from a terminal on your computer:

For all your chats (recommended): auto mode

hermes elm auto on

Every Desktop, TUI and CLI chat in this profile then has ELM, and a note saved in one chat is
remembered in all of them, like a normal memory. Restart Hermes if a chat is already open.
Never included: messaging apps (Telegram, Discord…), cron jobs, subagents and editor-connected (ACP) chats.

In a chatEffect
/elm privateThis chat stops reading shared notes; new notes and routing history from here stay in this chat. /elm private off undoes it.
/elm off / /elm onSwitch ELM off or back on for this chat only.
/elm statusShows auto_mode and shared_memory for this chat.

hermes elm auto shows the setting; hermes elm auto on --private keeps notes per chat instead of
sharing them; hermes elm auto off turns off every chat that was on only because of auto mode
(chats you granted yourself keep their grant; saved notes are kept).

For one chat at a time

Desktop (and TUI chats served by the Desktop/gateway): Hermes cannot tell these apps apart, so you grant access from a terminal on the same computer. The model can never grant access.

  1. Open the chat and send any message so it is saved.
  2. In a terminal run:
    hermes elm grant
    
    Pick the chat from the list (newest first) and confirm with y. For a named profile: hermes -p NAME elm grant.
    Desktop's Settings → Plugins → ELM page also shows the exact command for the chat you are in.
  3. Back in the chat, send /elm status, or press Refresh on the ELM settings page.

A grant lasts 24 hours by default (--hours up to 168). Turn it off with /elm off or hermes elm revoke; your saved notes are kept.

Classic CLI (hermes chat): just type /elm on in the chat.

Use it

You want to…Do this
Save something/elm save reference The staging server is at 10.0.0.5 (kinds: user, project, reference, skill)
Find something/elm search staging server, or just ask — matching notes are recalled automatically
Review model proposals/elm pending, then /elm approve <id> or /elm trash <id>
Browse / edit/elm list, /elm edit <id> <new text>, /elm restore <id>, or the memory manager in Settings → Plugins → ELM
Share with other chatsAuto mode shares by default; otherwise grant with --project NAME (same project) or --user (your shared user facts)
Keep one chat to itself/elm private
Keep a record of the conversationGrant with --capture (saves finished turns as plain text on this computer)
See every command/elm help

Model choices for the optional search planner and organizer: Settings → Models → Auxiliary → ELM Search / ELM Organizer.

Optional: memory-aware agent routing

When Hermes hands work to a subagent (delegate_task), ELM can choose which agent does it and
learn from how it went. You name your agents; each is a model or one of your Hermes profiles
(a profile brings its own model, skills and SOUL.md; subagents never read ELM memory):

# config.yaml of the profile (or: hermes config set delegation.backends.elm.agents.quick.model ...)
delegation:
  backend: elm
  backends:
    elm:
      default: quick          # used when ELM has nothing better
      agents:
        quick: {description: "Small, fast tasks", provider: deepseek, model: deepseek-v4-flash}
        deep:  {description: "Hard reasoning", provider: deepseek, model: deepseek-v4-flash, reasoning_effort: high}
        coder: {description: "Code changes in my repos", profile: coder}

Restart Hermes (or /new). The model now sees your agent names and descriptions and may pick one
per task. If it doesn't, ELM picks the agent that succeeded most on similar earlier tasks, else your
default, else the chat's own model. hermes elm routes shows your agents and how each did.

Routing learns only in chats where ELM is on, and like notes it stays private: a chat learns from
its own tasks, plus chats of the same granted project, plus chats sharing user memory when it does.
Chats without ELM just use your choice or the default. What ELM remembers: the agent's name, up to
12 plain words from the task (no numbers, codes or keys), whether it finished, and how long it took.
Never the task text or results. Revoking ELM stops recording; deleting a chat deletes its routing
history. Set learn: false to only use your choice or the default.
The ELM plugin must be enabled for routing (hermes plugins enable elm). Nous's official
Advanced Delegation plugin is an alternative backend without memory; only one backend is active.

What happens when…

  • A long chat is compressed: ELM keeps the chat's access and its private notes, even if Hermes moves the chat to a new session ID.
  • You delete a chat: its private notes and captured text are deleted too. Project and shared notes stay. Archiving a chat keeps everything.
  • You say "ok" or "thanks": nothing is recalled. Recall only uses meaningful words and doesn't repeat a note it already added to the chat.
  • You uninstall ELM: your data stays in <profile>/plugin-data/elm/ until you delete that folder.

Privacy and limits

  • This is personal memory for one person per Hermes profile, not a multi-user permission system. Use separate profiles for separate people.
  • Any app connected to the same granted chat (for example an editor client) can use that chat's memory.
  • Notes recalled into a chat are sent to your model provider like any other message.
  • Captured text is stored as plain text; don't enable capture in chats that contain passwords or secrets.
  • Not included yet: native chat folders, "important message" markers, semantic (embedding) search.

Uninstall

hermes plugins disable elm
hermes plugins remove elm

Then delete <profile>/plugin-data/elm/ if you also want to delete your memory (hermes elm doctor shows the profile path before you remove it).

More

README from the reviewed commit f62fc29; it updates when the author re-pins.

Developer Mode

Explore built-in Hermes toolsets and actions.

HermesAgent

Developer?

Create a Hermes extension for Desktop or agent workflows.

Developer Section