Dream

Dream: Memory Consolidation + Second Brain Sleeptime

You are performing a dream — a reflective pass over your memory files AND the Second Brain database. Synthesize what you've learned recently into durable, well-organized memories so that future sessions can orient quickly.

Memory directory: ~/.claude/projects/C--Users-<твой-логин>/memory/ (имя папки Claude Code кодирует из пути проекта — посмотри ls ~/.claude/projects/)

Second Brain — отдельный компонент, в пак он НЕ входит. Это локальная векторная база памяти со своим CLI (~/.brain/brain_sleeptime.py) и MCP-сервером second-brain. Проверить одной командой: ls ~/.brain/brain_sleeptime.py. Нет — пропускай Фазы 0, 6, 7 целиком: файловая консолидация (Фазы 1–5) самодостаточна и это законный полный прогон dream. В отчёте так и напиши: «Second Brain не установлен, фазы 0/6/7 пропущены».

Full v2 regulation (frontmatter schema, supersede protocol, decay/prune/promote formulas, TOP-INSTINCTS block format): references/memory-v2.md. Read it before Phases 2–5.

Related pattern (reference only, not installed): references/pi-llm-wiki-layered-pattern.md — 4-layer wiki architecture (immutable source packets / editable pages / auto metadata / config) mapped against our own bi-temporal system; flags one low-cost addition worth considering (append-only events.jsonl audit log for dream's own actions).

Ground Rules (apply to every phase)

• Anti-churn: if a file needs no change — do NOT touch it. No cosmetic edits, no timestamp refresh, no bulk frontmatter migration. Needless diffs break prompt-cache. "Memory already tight" is a valid outcome.
• Bi-temporal immutability: contradicted facts are never deleted or rewritten. New note with supersedes:, old note gets status: superseded / superseded_by: / invalid_at:. Only the index (MEMORY.md) swaps pointers freely.
• Budgets: MEMORY.md ≤200 lines and ≤25KB; index lines ≤200 chars.
• Authority criterion (consolidation): знание = «проверенное» только если (1) проверено реальным прогоном, (2) названа неработающая альтернатива, (3) описана конкретная решаемая проблема; иначе — status: pending, не канон.
• On demand only: dream runs when asked ("dream", "консолидируй память") or voluntarily once every few sessions. Not a cron, not a hook. Every phase is idempotent — a re-run is always safe.

Phase 0 — Brain Stats (Before) — только если Second Brain установлен

Capture brain state BEFORE consolidation so we can report the delta.

1. Call brain_stats MCP tool (server: second-brain)
2. Note: total memories, contacts, chunks pending embed, active commitments, awaiting items
3. Save these numbers for the "before/after" report in Phase 7

Phase 1 — Orient

• ls the memory directory to see what already exists
• Read MEMORY.md to understand the current index
• Skim existing topic files so you improve them rather than creating duplicates

Phase 2 — Rethink (Gather Signal + Detect Drift)

Look for new information worth persisting and old information that drifted. Sources in priority order:

1. Drifted memories — notes that contradict what the codebase / configs / reality show now
2. Contradictions between notes — two files that disagree with each other
3. Stale entries — memories about projects/tools that no longer exist or changed
4. Oversized entries — MEMORY.md lines over ~150 chars that carry content belonging in topic files

Don't exhaustively read everything. Look only for things you already suspect matter.

For every conflict found — mark it for the supersede protocol in Phase 3. Do NOT resolve conflicts by deleting or rewriting the older note.

Phase 2b — Transcript pass (по запросу / когда заметки отстали)

Фазы 1-2 смотрят только на то, что УЖЕ записано. Отдельный слой сырья — транскрипты прошлых сессий: там лежат уроки, которые в заметки не попали вовсе. Запускай этот проход, когда dream вызван с фокусом («dream по проекту X», «что я вынес за неделю») или когда сессий прошло много, а новых заметок почти нет.

• Фокус — параметр прохода. Тот же корпус под разным фокусом раскладывается по-разному; это ожидаемо. Фиксируй формулировку фокуса в отчёте, иначе потом не понять, почему память организована именно так. Без фокуса — обычный полный проход по Фазам 1-2.
• Не грузить транскрипты целиком. Работай точечно: search_chats.py search "<тема>" --days N → timeline <id> → get <id,id> только по отобранным якорям. Тот же принцип, что и с любыми большими данными: сузить, потом читать.
• Извлекать только durable: root cause + фикс, решения владельца («выбрал X потому что Y»), неочевидное поведение инструментов/библиотек, рабочие конфиги, сработавшие неочевидные подходы. Ход задачи, промежуточные попытки и одноразовые детали — не память.
• Извлечённое из транскрипта не канон по умолчанию: пока не выполнен authority criterion (см. Ground Rules) — status: pending. Ссылку на сессию-источник клади в evidence.
• Дальше извлечённое идёт обычным путём: Фаза 3 (merge/supersede), Фаза 4 (decay/prune/promote).

Кандидатный слой для крупной реорганизации. Если проход переписывает не пару файлов, а перекладывает структуру памяти — пиши сначала в _candidate/, сравни с живыми файлами, и только после просмотра принимай (принятое переносится, отклонённый кандидат удаляется целиком). Для точечных правок это лишняя церемония — там действует anti-churn.

Phase 3 — Consolidate (with supersede protocol)

For each thing worth updating:

• Merge new signal into existing topic files rather than creating near-duplicates
• Convert relative dates ("yesterday", "last week") to absolute dates
• Contradicted facts → supersede, don't delete: create a new note (id, valid_at, supersedes: [old-id]); in the old note set only status: superseded, superseded_by:, invalid_at: — body stays intact. Trivial typos/paths may be fixed in place without ceremony.
• Frontmatter v2 upgrade on touch: any note you are editing anyway gets upgraded to the v2 schema (id, type, status, trigger, action, confidence, evidence, valid_at/invalid_at, last_confirmed, discovery_tokens — see references/memory-v2.md §1). Never migrate untouched notes in bulk.
• Confirmations: when a session proved a note right — append a dated line to evidence, set last_confirmed, optionally +0.05 confidence (cap 0.98)

Phase 4 — Decay, Prune, Promote

Apply the lifecycle math from references/memory-v2.md §3–5:

• Decay: effective_confidence = confidence − 0.02 × weeks_since(last_confirmed || valid_at), floor 0.05. Compute in-flight; persist into a file only if you're writing it anyway or the value crosses an action threshold (anti-churn).
• Prune: status: pending + older than 30 days + effective confidence < 0.3 → move file to _archive/, drop its index line. Never prune superseded notes (they are history) or live references.
• Promote: a project note confirmed in ≥3 distinct sessions AND genuinely cross-project → copy to global memory dir as user/feedback, index it globally, mark the project copy status: promoted.

Phase 5 — TOP-INSTINCTS + Index

Rebuild TOP-INSTINCTS (no hook — dream itself does this every pass):

1. Collect type: feedback, status: active notes; rank by effective (decayed) confidence, tie-break by newer last_confirmed; take top K=7 (5–10 ok)
2. Rewrite ONLY the block between <!-- TOP-INSTINCTS:BEGIN --> and <!-- TOP-INSTINCTS:END --> markers in MEMORY.md (insert the marked block after the intro if missing). Line format: N. `conf` trigger → action — [id](file.md). Exact format and idempotency rules: references/memory-v2.md §6
3. Anti-churn: if the rendered list is unchanged except the date stamp — don't write

Then update the rest of MEMORY.md so it stays under 200 lines AND ~25KB:

• Swap index pointers from superseded notes to their successors; remove pointers to archived notes
• Demote verbose entries: if an index line is over ~200 chars, shorten it, move detail to topic file
• Add pointers to newly important memories
• Group related entries together (by topic, not chronologically)
• The hand-curated RULES — ALWAYS APPLY section stays hand-curated — TOP-INSTINCTS complements it, doesn't replace it

Rebuild memory graph (Layer 1 — after note/link/supersede changes settle): python ~/.claude/scripts/memory_graph.py build, then dangling + orphans — dangling [[links]] (target has no note) and high-degree orphans are hygiene signals: create the missing note, fix a name mismatch, or link the orphan. The graph (~/.claude/memory-graph/graph.db) powers recall routing «что связано с / хронология / что заместило».

Phase 6 — Second Brain Sleeptime — только если Second Brain установлен

After file-based consolidation, run the Second Brain sleeptime pipeline. Нет ~/.brain/brain_sleeptime.py — фаза пропускается, это не ошибка.

Run sleeptime via CLI:

python ${HOME}/.brain/brain_sleeptime.py run

This executes the full consolidation pipeline:

1. SHMR pass — Self-Harmonizing Memory Review: clusters similar memories by cosine similarity, keeps best version, dampens duplicates
2. Forgetting pass — Weibull decay on low-importance, unused memories
3. Contradiction scan — checks recent 7-day memories for contradictions
4. Pheromone decay — decays entity importance scores in the knowledge graph
5. Contact decay — recalculates relationship_strength for all contacts
6. Stats snapshot — logs final state to action_log

Then generate insights:

python ${HOME}/.brain/brain_sleeptime.py insights

This produces data-driven insights: hot entities, frequent memories, contact patterns, memory growth, pending vectorization.

If CLI fails (e.g. missing deps), explain:

Second Brain sleeptime can be run manually:

cd ${HOME}/.brain
python brain_sleeptime.py run
python brain_sleeptime.py insights

Phase 7 — Brain Stats (After) + Report

Без Second Brain отчёт состоит из одной секции «File Memory (Phases 1-5)», а секции «Second Brain Sleeptime» и «Brain Stats Delta» опускаются.

1. Call brain_stats MCP tool again to capture post-consolidation state
2. Compare before/after numbers
3. Report the combined results in this format:

Dream Report

File Memory (Phases 1-5):

• Фокус прохода (если задавался) + сколько сессий-транскриптов просмотрено в Фазе 2b и сколько уроков из них извлечено (или «transcript pass не запускался»)
• Files updated / created / superseded / archived / promoted
• TOP-INSTINCTS: rebuilt or unchanged (list the K entries if rebuilt)
• MEMORY.md line count and size

Second Brain Sleeptime (Phase 6):

• SHMR: N clusters found, N memories harmonized/dampened
• Forgetting: N memories archived/dampened by Weibull decay
• Contradictions: N found among recent memories (list top 3 if any)
• Pheromone decay: N entities decayed
• Contact decay: N contacts recalculated
• Insights generated: (list each insight)

Brain Stats Delta:

Output

Return the full Dream Report above. If file memories were already tight, say so — that is a valid outcome (anti-churn). If brain sleeptime had errors in any step, report which steps succeeded and which failed.