Learning Notebook¶
The Notebook lets you teach a bounded lesson, review its sources, and decide whether an agent should use it later. Start with the complete report-format exercise.
Where to find it¶
Web: Brain Mem → Lessons → choose an agent. iPhone: Agents → choose an agent → Learning. Enable Learning Notebook in the saved agent first. Learning-related reads and writes require authorized access.
The lifecycle¶
- Draft: explicit Teach input or an enabled agent's proposal creates a candidate, with source quotations. No candidate means nothing was saved.
- Review: inspect the actual content, source, type, and limitations.
- Approve: only approved guidance becomes eligible for future use.
- Test: start a fresh chat with the same agent/owner; inspect the result.
- Disable or refine: remove unhelpful guidance from future runs. Restoration is another draft and another review—not instant reactivation.
Preference/fact guidance can be supplied to a future run; procedural skills can be retrieved on demand. Retrieval is bounded, not a promise to load every lesson into every prompt. Search is weighted keyword matching, so wording matters.
Which kind of memory should I use?¶
| Need | Mechanism | Important distinction |
|---|---|---|
| “Remember the format I prefer” | Notebook preference | Explicit user preference, reviewed before use |
| “Reuse a verified procedure” | Notebook skill | A bounded procedure with evidence and pitfalls, not arbitrary installed code |
| “Search these manuals” | Knowledge base | Source documents can be updated and retrieved; not permanent personal policy |
| “Continue this conversation” | Chat session | Old conversation context can still contain a disabled lesson |
| “Inspect broader stored memory/rulebooks” | Agent memory | Separate storage and review mechanisms |
| “Learn from Studio repairs” | Studio learning | Workflow authoring/repair patterns, not this private Notebook |
Controls that are easy to misunderstand¶
- Propose useful lessons: permits proposals during suitable runs. It never means automatically approve them. Off still allows explicit Teach/review.
- Loaded count: the guidance was made available; it is not an effectiveness score.
- Helpful / Needs improvement: feedback for that lesson, not a self-modifying prompt.
- Disable lesson: excludes future reads/runs; cannot erase already-sent context.
- Draft restoration: creates a reviewable version; does not grant new tool access.
Privacy and limits¶
Private lessons are scoped to an agent and authenticated owner. Sharing one key shares an identity; issuing a different managed key may create a different Notebook scope. Do not treat this as personal isolation between users sharing an admin key.
The local database is permission-restricted but not encrypted by this feature. Approved content may be sent to the agent's configured model. Do not store credentials or assume cloud inference stays on your server. A source quotation proves provenance, not factual truth; facts may become stale.
For exact YAML/API fields, retention, source eligibility, and the boundaries of the Hermes-inspired learning cycle, see the technical reference.