Lightweight Learning Loop

AgenticLocal includes a local Hermes-style learning loop:

observe -> reflect -> distill -> approve -> reuse -> refine

The first version is conservative by design. It stores completed-run experiences in SQLite, proposes draft memories or procedural skills, and waits for explicit user approval before reusing them. It does not fine-tune a model, call a vector database, or auto-activate generated skills.

Modes

Learning is on in draft mode by default:

python3 -m agentic_loop chat --learning draft
python3 -m agentic_loop serve --learning draft

--learning auto is accepted for forward compatibility, but in this MVP it keeps generated skills in draft status to preserve the approval boundary.

Disable it for a run or service:

python3 -m agentic_loop --learning off "Inspect data/sample.csv"
python3 -m agentic_loop serve --learning off

The procedural-skill threshold defaults to 2, meaning a similar successful tool pattern must recur twice before the runtime proposes a skill draft. Lower it for local experiments:

python3 -m agentic_loop chat --learning-threshold 1

Review Commands

Inside agentic-loop chat:

/learn
/learn approve ID
/learn reject ID
/skills
/skill KEY
/skill archive KEY

Approving a memory draft writes it to long-term memory with source learning. Approving a skill draft promotes it to an active local Markdown skill. Archiving a skill disables future reuse without deleting history.

HTTP API

Served mode exposes the same review surface:

GET  /learning/drafts
POST /learning/drafts/approve
POST /learning/drafts/reject
GET  /skills
GET  /skills/{key}
POST /skills/archive

Approval and rejection requests accept id or draft_id. Skill archive requests accept key or skill.

Storage

The SQLite database stores:

• experiences: one sanitized summary per completed run.
• learning_drafts: proposed memory, skill, or anti-pattern artifacts.
• skills: active or archived procedural skills.
• skill_uses: when matched skills were injected and how the run ended.

Events such as experience_recorded, learning_draft_created, skill_approved, and skill_used flow through the existing event timeline and SSE endpoint.

Skill Format

Skills are plain Markdown with small front matter:

---
key: release-checklist
title: Release Checklist
description: How to prepare an AgenticLocal release safely.
triggers:
  - release
  - changelog
  - push
status: active
---

## When To Use
...

## Procedure
...

## Pitfalls
...

## Verification
...

Only compact metadata is considered by default. Full Markdown is injected into the model context only when lexical trigger matching finds a relevant active skill for the current goal.

Safety

• Secrets, API keys, temporary credentials, and full command output are not stored by default.
• Generated skills stay in draft status until approval.
• Learning failures are non-blocking and emit learning_error events instead of failing the user task.
• Vector memory is intentionally deferred; the MVP uses lightweight lexical matching so it remains dependency-free and easy to inspect.