learn

Close out the session in one pass:

1. Audit — dispatch parallel lens audits against files changed this session; print the composite verdict
2. Consolidate — encode novel learnings as atomic memories; update existing over creating new; prune stale
3. Reconcile — bring STATUS.md / ISSUES.md / PLAN.md / README.md (and academic manifests if applicable) into alignment with current reality
4. Report — one compact block; no narration

Most sessions produce zero memories worth keeping and a handful of doc updates. That's correct — the work is in the files.

- **One idea, one memory.** Compound learnings split into separate files. (Minimum Information — Wozniak) - **Novel only.** Nothing derivable from code, git, docs, or CLAUDE.md. Memory is for the unwritten. (Derivability Test) - **Update before create.** New learning relating to an existing memory reconsolidates it — retrieval strengthens encoding. (Nader et al.) - **No ephemera.** 30-day test: would this change my approach in a month? If not, let it decay. (Ebbinghaus) - **Emotional salience first.** Corrections, errors, surprises, and confirmed non-obvious choices are the encoding priority. (LaBar & Cabeza) - **Interference prevention.** Merge near-duplicate memories; a cluttered system retrieves badly. (McGeoch) - **Documents decay too.** Stale STATUS/PLAN entries create proactive interference for the next session. Reconcile or let the doc stay silent. Run the audit first when available — it surfaces findings that become encoding signals and ISSUES.md entries.

Locate the audit dispatch reference. Check these paths in order, stop at the first hit:

1. <repo-root>/.claude/skills/_audit-shared/audit-dispatch.md (repo-local shared audits)
2. $CLAUDE_AUDIT_DISPATCH environment variable (user-configured absolute path)
3. references/audit-dispatch.md relative to this SKILL.md (bundled fallback, if shipped)

If found: follow it end-to-end: scope → classify → parallel dispatch → synthesize → save composite. Keep the composite inline.

If not found: print one line — "Audit skipped — no audit dispatch reference configured. Proceeding to memory consolidation." — and continue. This is the normal case for installs without the Opus Vita audit suite.

If scope is empty (no file changes this session): skip the audit dispatch but still proceed to encode — conversation-only learnings (user corrections, preferences, decisions) matter too.

Scan the full conversation (plus audit output, if one was produced) for learning signals. Priority order:

1. User corrections ("no", "don't", "stop doing X")
2. Errors or failures that revealed something non-obvious
3. User confirmations of non-obvious choices ("yes exactly", silent acceptance)
4. Architecture decisions with rationale
5. Audit findings tagged as surprising, cross-lens, or "wins worth noting"
6. New external resources or references
7. User preferences, constraints, stakeholder context

For each candidate, apply the atomization rules (load references/atomization-rules.md) in order:

1. "And" test — split compound learnings
2. 30-day test — discard ephemera
3. Derivability test — discard what code/git/docs already say
4. Specificity gradient — transferable principle, not line-42-specific
5. Type assignment — feedback / project / user / reference

For feedback and project types, structure as: rule/fact → Why: → How to apply:

Internal only — do not output this step.

Integrate with existing memory.

Locate the memory directory. Check in order, stop at first hit:

1. $CLAUDE_MEMORY_DIR environment variable
2. A memory: path specified in a user or project CLAUDE.md
3. Claude Code's default project memory dir — typically ~/.claude/projects/<cwd-slug>/memory/ where <cwd-slug> encodes the project path
4. <repo-root>/.claude/memory/ (repo-local fallback)

Create the directory if missing (empty MEMORY.md, then proceed).

1. Read MEMORY.md in the resolved memory directory
2. For each encoded candidate, search for topical overlap with existing memories
3. Overlap → reconsolidate. Read the existing file. Update it with new context. Rewrite. (Piaget — schema assimilation)
4. No overlap → create. Write a new file: {type}_{topic}.md with frontmatter (name, description, type). Add one-line pointer to MEMORY.md (under 150 chars)
5. While reading memories, prune any that contradict current code, have passed deadlines, or duplicate another. Remove stale, update MEMORY.md

Elevation check: note (don't act) if you see:

• 3+ feedback memories on the same mistake category → candidate hook
• 3+ feedback memories on the same workflow preference → candidate CLAUDE.md convention
• 3+ reference memories on the same system → candidate consolidated doc
• Multi-step process captured across memories → candidate skill

Load references/memory-science.md only if ambiguous about type assignment or specificity gradient.

Synchronize project documents with current reality. Load [references/document-reconciliation.md](references/document-reconciliation.md) only if creating a doc from scratch or a structural question arises.

STATUS.md — add session entry to "Recent Sessions" (date + one-line); update "Current State" counts; shift "In Progress" / "Upcoming"; clear or add blockers; prune past 10 sessions; cross-validate issue counts against ISSUES.md.

ISSUES.md — add new issues (P0–P3, unique IDs); move resolved to "Recently Resolved" (date + notes); promote 🔴/🟠 audit findings that aren't tracked; merge near-duplicates; prune past 10 resolved.

PLAN.md — check off completed tasks ([ ] → [x]); update phase status; add newly discovered tasks; collapse completed phases to one-line summaries; remove tasks no longer relevant.

README.md — update only if structural changes occurred (new deps, changed setup, renamed keys, restructured dirs). README serves newcomers, not session continuity.

Academic manifests (conditional) — if the session touched 1 PROJECTS/*/:

• Update _manifest.md: deliverable status + last_modified
• Reflect section state changes (draft → locked, etc.)
• Append handoff entries (date, recipient, commit hash) if sent to committee
• Orphan check: fold temp files into canonical or move to working/ with ISO timestamp
• Update frontmatter last_modified on edited _*.md files
• Append to _session-log.md if present

Create missing documents only when warranted: STATUS.md if code changed but none exists; ISSUES.md if issues discovered and none exists. Do NOT create PLAN.md or README.md unprompted.

Skip documents that need no changes. Reconciliation without change is noise.

Prune old audit reports (if a docs/audits/ directory exists). Keep last 20 composite reports and 60 days of per-lens reports. Monotonic growth here creates proactive interference when docs/audits/ accumulates hundreds of files — search, git, and future audits all suffer.

One compact block. No narration, no restating audit content (the audit was already printed inline).

Audit: [Ship it | Ship with fixes | Hold] — [one-line truth]
Memory: [N created, N reconsolidated, N pruned]  (list titles, one line each — omit section if N=0)
Docs: [list of files updated, one line each]     (omit section if none)
Elevation: [pattern detected — consider skill|hook|convention]  (omit if none)

If nothing worth persisting and no documents changed: Nothing to consolidate.

Load only when needed: - [references/memory-science.md](references/memory-science.md) — encoding principles (ambiguous type or specificity) - [references/atomization-rules.md](references/atomization-rules.md) — always loaded in encode step - [references/document-reconciliation.md](references/document-reconciliation.md) — structural doc questions