Skip to content

Local-first 🧠 memory for Hermes Agent that lives in your Obsidian vault and remembers project context. Nightly 😴 dream passes turn repeat corrections into confirmed preferences with measurable confidence. Adapters ship for Claude Code, Codex, and OpenClaw, with an MCP server for anything else.

TypeScript
3.1 KB
Raw Blame

Label suggestions for brain_labels

Use labels of the optional decision model (see ../decision-models.md for configuration, privacy and accounting). Controlled-vocabulary labels are declared per dimension in the schema pack (schema.labels) and assigned explicitly with assign, which stays fail-closed. suggest classifies a note over the declared vocabulary and never assigns anything.

Request

brain_labels operation suggest with the note path and optional dimensions (default every declared dimension), or o2b brain label <path> --suggest [--dimensions a,b].

One request per note. The state is the note's title (frontmatter title, else the file name, clipped to 200 characters) and body (private regions stripped, clipped to 4000 characters); the adapter redacts the whole body. Each requested dimension is one choice question dim_<i> over its declared values plus none (named _none when the vocabulary itself declares a value none, so the two stay apart in probabilities). The schema pack declares values without descriptions, so the options carry the value names only.

  • A note whose visibility carries private is refused with an error and nothing is sent. Through MCP a page reserved against remote reads answers as absent, exactly as show does.
  • An unknown dimension is refused as assign refuses it.
  • A dimension with more than 254 values is skipped with reason: "too_many_values" (a question carries at most 255 options, none included).

Result

Per dimension: { dimension, suggestion, probabilities, confidence, current, model, calibrated }, with current the value the note carries now (or null) and skipped listing the dimensions that were not asked.

  • off or no active configuration: { available: false, reason: "decision_model_off" }. This is not an error, and nothing is sent.
  • shadow: mode: "shadow"; the request is sent and recorded, and every suggestion, probabilities and confidence is null, so the host is not steered before the use has been evaluated.
  • enforce: the chosen value; none, or a confidence below 0.6, gives suggestion: null.
  • Any failure: { available: false, reason: <degrade reason> }.

Suggest before assign: read the suggestion, then assign explicitly with assign if it fits.

Accounting and evaluation

The request writes one decision_model_call record with note_path and, per dimension, the current value, the chosen option, the value that would be suggested (in both modes), its confidence and probability. No note text is recorded.

The report (reports/labels.ts) measures acceptance as is: it compares each latest suggestion per note and dimension with what the note carries now. assign writes only the note's frontmatter, so the note is the record of later assignments. accepted means the dimension now holds the suggested value, overridden that it changed to something else, unchanged that nothing was assigned since, already that the suggestion equalled the value the note had. Acceptance is accepted over accepted plus overridden, reported per mode, so shadow data can be evaluated before enforce.