git:20260813.53fd4ac to git:20260902.8ca5fc4

31 added, 24 removed. Audit A to A.

---

name: bosskuai-continuous-learning

description: Use this after meaningful tasks, reviews, incidents, or repeated observations to decide the smallest durable artifact that should capture the lesson and keep future sessions from relearning it.

---

# BosskuAI Continuous Learning

Use this skill when recent work produced a lesson worth keeping beyond the current chat.

+ ## How this differs from nearby skills

+ - **`bosskuai-permanent-memory-orchestration`**: the mechanics of `.bossku/memory/`, `bossku remember`, and the Obsidian export; this skill decides what deserves to go there.

+ - **`bosskuai-rules-distill`**: promotes principles seen in 3+ places into rules; this skill handles a single fresh lesson.

+ - **`bosskuai-skill-creator`**: builds or revises a skill once this skill decides a skill is the right artifact.

## Meaningfulness gate

- Meaningful if any of these are true: files changed, a decision was made, a bug/risk was found, a pattern repeated, or a workflow improvement emerged.

- Trivial if all are false. In that case, stop after noting there was no durable delta.

- ## Routing ladder

+ ## Routing ladder (smallest artifact that changes future behavior)

- - `plan-log.md`: compact pre-execution plan worth retrieving before or during execution

- - `learning-log.md`: default chronological handoff

- - `agent-profile.md` / `project-understanding.md`: durable workspace facts

- - `bug-patterns.md`: repeated defect class or high-severity near miss

- - `market-notes.md`: durable GTM or competitor lesson

- - `instincts/`: atomic behavioral lesson with a confidence score (see below)

- - checklist / pitfall / playbook / skill / rule: when the lesson should change future behavior beyond one repo note

+ | Lesson looks like | Artifact |

+ |---|---|

+ | a plan another session may continue | `bossku remember --kind plan` → `.bossku/memory/plans.md` |

+ | a verified outcome, recurring bug, or verification result | `bossku remember --kind learning` → `.bossku/memory/learnings.md` |

+ | a choice worth defending later | `bossku remember --kind decision` → `.bossku/memory/decisions.md` |

+ | a durable fact about the repo, stack, or users | `bossku remember --kind project` → `.bossku/memory/project.md` |

+ | a small repeatable behavior with uncertain confidence | instinct file in `.bossku/memory/instincts/` (below) |

+ | a repeatable verification step, recurring trap, or reusable procedure | checklist, pitfall, or playbook under BosskuAI `references/` |

+ | a change to how the assistant works across many tasks | skill (`bosskuai-skill-creator`) or rule (`bosskuai-rules-distill`) |

+ `bossku remember` redacts secrets and exports the four kind files to the Obsidian vault when one is configured, so a memory write is also the Obsidian sync.

## Instinct model (confidence-weighted micro-lessons)

When a lesson is a small repeatable behavior rather than a fact or a log entry, store it as an **instinct**: one trigger, one action, confidence-weighted. Adapted from ECC continuous-learning-v2.

One file per instinct in `.bossku/memory/instincts/`, named by id:

```yaml

---

- id: prefer-runtime-core-blocks

- trigger: "when adding or editing a pipeline agent persona"

- confidence: 0.7 # 0.3 tentative → 0.9 near certain

+ id: rebuild-index-after-skill-edit

+ trigger: "when a SKILL.md name, description, or frontmatter changes"

+ confidence: 0.8 # 0.3 tentative → 0.9 near certain

scope: project # project (default) or global (true in 2+ projects)

- evidence: "2026-05-29 persona token audit: full-file injection cost 6,801 tokens; runtime-core cut it 79%"

+ evidence: "2026-08-18 validate failed on index fingerprint drift after a description edit"

---

- Action: include a  block; verify with bosskuai:sync-personas --dry-run.

+ Action: run `python -m bossku skills index --root .` then `python -m bossku validate --root .` before declaring done.

```

Rules:

- **Atomic** — one trigger, one action; split anything compound.

- **Confidence moves on evidence** — raise by ~0.1 when the instinct is confirmed again, lower when contradicted; delete below 0.3.

- **Promote at high confidence** — an instinct that stays ≥0.8 and keeps firing should graduate up the ladder into a checklist, rule, or skill section; record the promotion and remove the instinct.

- **Scope before sharing** — keep instincts project-scoped by default; mark `scope: global` only when the same behavior proved out in 2+ repos.

- - **Runtime parallel** — the in-app equivalent is `LearningEngine` pattern/failure/preference events scored by `PostMemoryEvaluationService`; do not duplicate what the app already captures, but an app-side failure pattern may justify an editor-side instinct.

+ - Instincts are not exported to Obsidian; promote them if they need to travel.

## Workflow

1. Gather evidence from the finished task, diff, tests, review findings, and memory state.

2. Write each candidate learning in one sentence.

- 3. Choose the smallest artifact that will reliably change future behavior.

- 4. Keep pre-execution intent in `plan-log.md`; keep post-execution outcomes and lessons in `learning-log.md` when no stronger promotion is justified yet.

- 5. Fix stale or contradictory memory if discovered.

- 6. If indexed memory files changed, run `python3 ./ai-assistant/scripts/vector_memory.py sync`.

- 7. Record whether the learning was applied, deferred, or rejected.

+ 3. Choose the smallest artifact from the ladder that will reliably change future behavior.

+ 4. Write it: `bossku remember --project . --kind <kind> "<one to three lines>"` for memory kinds; edit the file directly for instincts and references.

+ 5. Fix stale or contradictory memory if discovered (append a dated correction).

+ 6. Record whether the learning was applied, deferred, or rejected, and what evidence would justify promotion later.

## Guardrails

- Do not promote vague advice.

- Do not store secrets, sensitive customer data, or transient debug chatter in memory.

- Do not create a new skill when a checklist, pitfall, or short memory note is enough.

- Do not turn one-off observations into universal rules.

+ - Do not write the same lesson into two artifacts; the weaker one should point at the stronger.

## Output

Return:

- signals reviewed

- candidate learnings

- chosen artifact for each learning

- - smallest safe update

+ - smallest safe update (and the `bossku remember` result, including export status)

- freshness issues found

- next promotion actions

## References

- `../../references/memory-first-handoff-protocol.md`

- `../../references/checklists/learning-promotion-checklist.md`

- `../../references/checklists/continuous-learning-checklist.md`

- `../../references/playbooks/continuous-learning-playbook.md`

- - `.bossku/memory/plans.md`

- - `.bossku/memory/learnings.md`

- - `.bossku/memory/project.md`

+ - `../../docs/memory.md`