Hermes Bots

From its catalog listing
Hermes Bots: a calmer Hermes Bot Chat that still does the work: short agentic replies, a block-only safety guard, a Desktop status card, mid-turn asks that are no longer dropped, background results delivered back to the same chat, a local short-term / long-term memory (decisions, project state, lessons; /memoria) with no API keys, and an opt-in panteón: Hermes may offer in one line to create a specialist bot and creates it only after your yes. Other sessions untouched.
Good to know before installing
in sessions titled "Bot Chat" (and their subagent children) it injects a reply-style instruction, a bounded context / memory block, pending-ask reminders and (Desktop half enabled) a one-line specialist offer; blocks tool calls matching its rules (secrets, deletes, destructive git, protected-branch pushes, outbound messages, expensive models, kanban-lock bypasses, obfuscated exec, memory-injection) and, when team.assignees is configured, rewrites kanban_create arguments to add serialization parents and the teammate assignment; wraps tool output that looks like a prompt injection in an "untrusted data" label; it never approves anything. It opens <HERMES_HOME>/state.db, <HERMES_HOME>/kanban.db and <HERMES_HOME>/cron/executions.db (plus cron output files) read-only; writes plugin-local files under <HERMES_HOME>/bot-chat-agent/ (memory.db, deliveries.db, status.json, panteon-auto.json, dm-delegations.json, optional project-card.md) and its own team.assignees plugin setting; registers /memoria, /botmemory, /panteon and /pantheon; serves a small dashboard API at /api/plugins/bot-chat-agent; the Desktop half (opt-in in Capabilities → Plugins) calls the app's own profiles.create / profiles.configure for ONE preset profile, only after the user's explicit yes; changes the Desktop reasoning.collapsedByDefault setting while a Bot Chat is focused and restores it on leave/disable; makes no third-party network calls, no telemetry and no credential reads. Background work uses Hermes kanban / the gateway dispatcher when team.assignees is configured; on Hermes 0.23, results are offered into the next turn. A ContextEngine named bot-chat is registered but off until context.engine: bot-chat is set.
From its catalog listing.
How to install
- Install it. Hermes asks for any API keys it needs, then asks whether to turn it on.
hermes plugins install bot-chat-agent - Said no? Turn it on later:
hermes plugins enable bot-chat-agent - Restart Hermes so it loads the plugin.
Some plugins also need a sign-in or settings: see Full setup & settings. Memory plugins ask whether to become your memory provider instead. Steps follow the official plugins guide (opens in a new tab).
- Category
- Memory & notes · Hermes Desktop add-ons
- Maintained by
- Tormen28 (Community)
- Needs
- No API keys declared
- Adds hooks
pre_llm_callpre_tool_callpost_tool_calltransform_tool_resultpost_llm_callagent_loop_stoppedkanban_task_completedkanban_task_blockedkanban_task_claimedon_kanban_worker_spawned- Works on
- linux, macos, windows
- Needs Hermes
- 0.22.5 or newer
- Version
- 0.3.0
- In the catalog
- 2026-11-07, updated 2026-11-11
Known issues (from its catalog listing)
- Per-turn cost: the plugin's own pre_llm_call hook takes about 15 ms p50 / 20 ms p95 on a Windows 11 PC and about 5 ms p50 on Linux (SQLite connections are shared per turn). Measured only with a mock model, so real-model time-to-first-token and the quality of the deterministic compactor are not measured; with context.engine: bot-chat requests were ~30% smaller than without the plugin in a mock 300-turn run. The engine is off by default.
- The panteón offer is an instruction injected into Bot Chat turns; a real model follows it most of the time, not guaranteed. The Desktop half must be switched on in Capabilities → Plugins for offers and profile creation.
- Background-result delivery on Hermes 0.23 uses the next-turn offer path; durable in-transcript writes need the core B1 API, which 0.23 does not expose. The core bot-loop guard (gateway.bot_loop_guard.max_hops, default 4) can refuse later delegations to the same teammate in one chat; the card then shows an error.
- Long-term memory is written only when the turn text contains clear cues (Decision:, Decisión:, We decided, Decidimos, Lesson:, Remember that, Recuerda que, From now on, …). There is no model and no network.
- Language follows the user (es/en/fr/pt/de, detected offline from the latest message; language: es|en|fr|pt|de forces one). Plugin-written text (status card, /memoria, /botmemory, /pantheon) is translated; lines the model writes (panteón offer, ready notice, stop explanation) get an instruction to answer in that language, which a real model usually but not always follows. Specialist role prompts and guard block messages stay en/es.
- hermes plugins remove takes out the Python half only; your data folder and the Desktop half (installed by the core) stay until you delete them.
Using the Hermes desktop app? See where plugins live

Full setup & settings ↗Source code ↗
More in memory & notes
New Hermes Agent updates every day on X. Follow @HermesWatcher