NeuralMind v4.9.0 — Hermes-Agent gets NeuralMind’s context in every turn, without a tool call

v4.8.0 gave Claude Code a session recap: a new session starts with where the last one left off. Claude Code already got NeuralMind’s related files with every prompt, through its hooks. Hermes-Agent got neither unless the agent decided to call a tool. v4.9.0 adds a Hermes plugin that gives Hermes both, before the model runs, with no tool call.

NeuralMind for Hermes-Agent

On Hermes, NeuralMind was something the agent had to decide to use: an MCP server, or a skill that runs the CLI. Each answer cost a tool call, and when the agent didn’t think to call it, nothing happened. v4.9.0 adds a Hermes plugin that puts NeuralMind’s context into every turn before the model runs, the way the hooks already do in Claude Code.

neuralmind build /path/to/project
neuralmind install-hermes-plugin /path/to/project

The command writes the plugin into the Hermes home’s plugins/neuralmind/: the one --hermes-home names, else the one plain hermes uses: the active Hermes profile’s, if hermes profile use selected one (<root>/profiles/<name>), else $HERMES_HOME, else Hermes’s default ~/.hermes (on Windows, %LOCALAPPDATA%\hermes). When that’s a profile, the output names it. The command then runs hermes plugins enable neuralmind in that same home, but only in a home Hermes has already set up (one with a config.yaml, .env or state.db); anywhere else it tells you to enable the plugin once Hermes is set up. Start a new Hermes session, or restart the gateway, to load it.

• Every turn (Hermes’s pre_llm_call hook): the files and recorded decisions related to the user’s message, the same block Claude Code’s UserPromptSubmit hook adds. Hermes appends it to that turn’s user message, not to the system prompt, so the prompt cache isn’t invalidated.
• A session’s first turn also starts with the session recap (v4.8.0). Hermes counts a turn as first only when the session has no earlier messages, so a resumed session doesn’t get it.
• After write_file and patch (Hermes’s post_tool_call hook): each file the edit wrote is recorded, by the absolute path Hermes reports (files_modified), on a background thread, for the recap and the synapse layer. Only an edit that landed counts: Hermes has to report its status as ok, so a failed, cancelled, timed-out or blocked edit isn’t recorded. A file a V4A patch deletes or moves away isn’t listed as edited, a patch that changes nothing isn’t recorded, and neither is a file changed through the terminal.
• Subagents and cron jobs are skipped. A subagent’s message is written by its parent agent, and a cron job’s (Hermes’s cron platform) is a scheduled prompt, not one you typed. Neither is recorded or answered with recall, neither runs the SessionStart action, and their edits aren’t recorded either.
• One record for both agents. Hermes and Claude Code write to the same .neuralmind/recaps/, so a Hermes session can start with what the last Claude Code session in the project did, and the other way round.

How it works: the plugin is a small stdlib-only file. Each action runs python -m neuralmind _hook <action> with the payload Claude Code would send, using the Python interpreter that ran the install. Hermes gets the same behavior and the same switches (NEURALMIND_BYPASS, NEURALMIND_SYNAPSE_INJECT, NEURALMIND_SESSION_RECAP …), and NeuralMind doesn’t have to be installed in Hermes’s own environment. The subprocess runs from the plugin’s own directory with PYTHONSAFEPATH=1, so a repository you work in can’t put its own neuralmind module ahead of the installed one.

Which project a Hermes session belongs to:

1. NEURALMIND_PROJECT, if set;
2. else the path given to install-hermes-plugin;
3. else Hermes’s terminal working directory (TERMINAL_CWD);
4. else, only when TERMINAL_CWD isn’t set, the directory Hermes runs in.

The first of these that has been built wins; if none has, the plugin does nothing. TERMINAL_CWD is read from the Hermes process’s environment, so a cd the agent runs later in a session doesn’t change the project. That’s the directory the terminal CLI and a standalone gateway work in. A gateway session (Telegram, Discord …) uses the gateway’s terminal working directory (terminal.cwd, else MESSAGING_CWD, else your home directory), which is rarely the project you mean, so pin one at install or set NEURALMIND_PROJECT. Hermes Desktop, ACP editor sessions and per-session workspaces keep each session’s directory elsewhere, not in TERMINAL_CWD, so with those, too, pin a project or set NEURALMIND_PROJECT. Whenever a gateway session resolves to a built project, pinned or not, every message in it is recorded for the recap, redacted like any other prompt; NEURALMIND_SESSION_RECAP=0 turns that off.

A pin applies to every Hermes session that uses that Hermes home, whatever directory it runs in: each one gets the pinned project’s context, and its prompts are recorded in that project. A session in another repository also puts the paths of the files it edits into the pinned project’s synapse store, from where neuralmind memory publish can carry them into the committed team-memory bundle. If you use Hermes across several projects, install without a path (or with --unpin, if you pinned one before), so the plugin follows the directory Hermes works in (in the terminal CLI and a standalone gateway, as above) and does nothing in a repository NeuralMind hasn’t built. Re-running the install without a path keeps an earlier pin; --unpin clears it.

Limits:

• One subprocess per turn (two on a session’s first turn: the recap, then recall), plus one per edited file. Each waits at most NEURALMIND_HERMES_TIMEOUT (default 8 seconds), so a first turn can wait up to twice that; one that times out or fails is left out, and the turn goes ahead with whatever context the others returned. Keep twice the timeout below Hermes’s plugins.hook_callback_timeout (default 30 seconds): a plugin that runs past it loses its whole block, and Hermes skips the plugin’s per-turn hook for the next 60 seconds.
• The context stays in Hermes’s session history. Hermes stores the turn’s message with NeuralMind’s block in it, so the block, recap included, is sent to your model provider again with that session’s later turns.
• On a session’s first turn the plugin runs NeuralMind’s whole SessionStart action, the same as Claude Code, including a synapse decay tick, the team-memory import, clearing the session-scoped (ephemeral) associations, and the SYNAPSE_MEMORY.md export (copied into Claude Code’s auto-memory directory when that exists; NEURALMIND_SYNAPSE_EXPORT=0 turns the export off).
• Tested against a Hermes v0.21.5 main-branch build (0.21.5+5355), by calling its plugin loader and hook dispatch directly, not yet in a live Hermes conversation. It relies on pre_llm_call accepting {"context": ...}, which that version documents.
• The plugin is a copy. pip install -U neuralmind doesn’t update it; re-run neuralmind install-hermes-plugin after upgrading. A re-run doesn’t turn back on a plugin you switched off with hermes plugins disable neuralmind: it says so and leaves it disabled.
• Not measured. As with the recap, we haven’t measured what the context changes in Hermes’s answers.

neuralmind install-hermes-plugin --uninstall disables it in that same home and removes it. If plugins/neuralmind is a symlink, it removes the link, never what it points to; the install refuses to write through one.

Also in this release

• ghcr.io/dfrostar/neuralmind:latest follows the newest release. The Docker workflow moved :latest to whichever version it built last, so rebuilding an older tag pointed it back at that version. On 2026-10-06 a re-pushed v1.16.0 tag pointed :latest at v1.16.0’s image. A build now pushes only its own :vX.Y.Z. A last step, one run at a time, then points :latest at the image of the highest vX.Y.Z tag, so neither an older rebuild nor two overlapping releases can move it backwards. This holds for tags cut from this release on. A tag push runs the workflow file at the tagged commit, so re-pushing an older tag still runs that tag’s old workflow and can still move :latest; only a tag ruleset that restricts who can push v* tags prevents that. To pin, pull :v4.9.0 rather than :latest.
• The session recap no longer depends on directory order. Two sessions whose last activity shares a timestamp (writes inside one clock tick, about 15.6 ms on Windows) were decided by whichever file the directory listed first. The recap now goes to the one modified last.

Per-agent expectations

Settings

The plugin runs the same hook actions as Claude Code’s hooks, so their switches apply to it too: NEURALMIND_BYPASS, NEURALMIND_SYNAPSE_INJECT, NEURALMIND_SESSION_RECAP, NEURALMIND_NO_LEARN and the rest.

Upgrading

pip install -U neuralmind, then run neuralmind install-hermes-plugin [path] once. The installed plugin is a copy, so run it again after later upgrades; a re-run keeps an earlier pin and leaves a plugin you disabled disabled.

Related

• Use case: Hermes-Agent with code memory in every turn
• Use case: Pick up where you left off
• Integration guide: Hermes-Agent
• CLI reference: install-hermes-plugin, Environment Variables
• Previous release: v4.8.1 · v4.8.0