remember-evolve

name	remember:evolve
description	Evolve the brain — consolidate entities, reflect on beliefs, promote top beliefs to Persona. Run weekly or on demand.

/remember:evolve — Evolve the Brain (periodic, weekly recommended)

The deeper LLM-driven evolution layer. The deterministic Phase 3 (promote) runs automatically after every capture — that keeps Persona.md ## Top Beliefs current in real time. This skill is what does the work that promote alone can't: re-synthesizing entity profiles, re-scoring belief confidence semantically, marking stale beliefs that should retire.

Without periodic evolution, the brain becomes append-only:

• People/, Projects/, Areas/ files accumulate log entries but never get a refreshed ## Profile overview
• Beliefs keep their initial confidence forever, even after 15 supporting captures that should push them to 0.95
• 6-month-old beliefs with no new evidence stay marked stable instead of stale

Three phases:

1. Consolidate (LLM) — re-synthesize entity profiles from accumulated facts and beliefs
2. Reflect (LLM) — re-evaluate beliefs against evidence, update confidence and freshness, mark contradictions
3. Promote (deterministic, no LLM) — also runs here, usually a no-op because auto-promote already kept Top Beliefs current

Recommended cadence: /loop 7d /remember:evolve (weekly) or /loop 30d /remember:evolve (monthly). Idempotent — safe to run any time.

After each Phase that writes (Phase 1 entity re-synthesis, Phase 2 belief frontmatter updates, Phase 3 Persona.md update via promote.js), call node ${CLAUDE_PLUGIN_ROOT}/scripts/schema.js validate <filepath> on the touched files. Surface its warnings (e.g. "confidence defaulted to 0.5 — review") in your final report.

---

Flags (parsed from user input)

Flag	Effect
--consolidate-only	Run Phase 1 only
--reflect-only	Run Phase 2 only
--promote-only	Run Phase 3 only (no LLM — pure script call)
--dry-run	Log intended changes; write nothing

If no flag → run all three phases in order.

---

Step 1: Resolve brain path & load config

1. Read $REMEMBER_BRAIN_PATH (fallback ~/remember). Call this {brain}.
2. If missing → tell user to run /remember:init and stop.
3. Run: node ${CLAUDE_PLUGIN_ROOT}/scripts/build-index.js --compact to get the brain index. Use throughout to resolve entities and prevent duplicates.
4. Load thresholds + auto_promote flag:

   node -e "const {loadEvolutionConfig} = require('${CLAUDE_PLUGIN_ROOT}/scripts/config'); console.log(JSON.stringify(loadEvolutionConfig()));"
   

   Defaults: promotion_confidence=0.85, promotion_sources=5, stale_days=90, consolidate_touches=5, top_beliefs_n=10.

Step 2: Read REMEMBER.md

Cascading: {brain}/REMEMBER.md (global) + ./REMEMBER.md (project, if cwd is in a project). User's ## Promotion Thresholds section overrides the config file values.

---

Phase 1 — Consolidate (LLM)

Skip if --reflect-only or --promote-only.

Goal: re-synthesize entity profile files (type: observation) for entities that crossed the consolidation threshold.

Step 1a: Find entities ripe for consolidation

For each People/<x>.md, Projects/<x>/<x>.md, Areas/<x>.md:

1. Parse frontmatter. Skip if missing type: observation.
2. Compute touches_since_last_consolidate — count of session entries / mentions in Journal/ that reference this entity (via [[wikilinks]]) since last_consolidated.
3. Include in candidate list if touches_since_last_consolidate >= config.thresholds.consolidate_touches (default 5).

Step 1b: For each candidate, re-synthesize

For each candidate file:

1. Read the entity file + all linked Notes/Decisions (via related: and grep for [[Notes/...]] mentioning the entity).
2. Synthesize an updated ## Profile (or ## Overview) section from accumulated facts + beliefs. Reference linked beliefs by wikilink, do not copy them.
3. Update sections (using Edit tool — surgical, never replace whole file):
   • ## Profile — synthesized summary
   • ## Active Facts — list of [[Notes/...]] world-facts about this entity
   • ## Active Beliefs — list of [[Notes/...]] beliefs with current confidence
4. Update frontmatter: last_consolidated: {today}, sources_count: {touches}, freshness: stable.
5. Append to evolution log:

   node ${CLAUDE_PLUGIN_ROOT}/scripts/evolution-log.js CONSOLIDATE "{path} touches=+{n}"
   

Never overwrite history sections (## Interactions, ## Log, ## Meetings). Only refresh synthesized sections.

Dry-run: report what would be consolidated, write nothing.

---

Phase 2 — Reflect (LLM)

Skip if --consolidate-only or --promote-only.

Goal: re-score beliefs against accumulated evidence and counter-evidence; update freshness trend.

See @reference.md for the full freshness state machine and confidence-update heuristics.

Step 2a: Find beliefs to reflect on

Walk Notes/*.md. Filter to type: belief. Skip beliefs reflected within last 7 days unless --reflect-only is explicitly invoked.

Step 2b: Per-belief re-evaluation

For each belief file:

1. Read frontmatter and body. Note current confidence, freshness, sources_count, evidence, counter_evidence.

2. Score new confidence based on:

   • Ratio of evidence count to (evidence + counter_evidence) count
   • Recency of latest evidence (penalty if all evidence is older than stale_days)
   • Strength of contradicting counter-evidence (if any explicit "user said the opposite" entries exist)

3. Update freshness using the state machine in @reference.md:

   • Confidence rose ≥ 0.1 → strengthening
   • Confidence dropped ≥ 0.1 OR new counter_evidence in this run → weakening
   • No new evidence in stale_days (default 90) → stale
   • counter_evidence.length > evidence.length → contradicted
   • Otherwise → stable

4. If freshness changed OR confidence changed by ≥ 0.05:

   • Update frontmatter via Edit tool (surgical)
   • Append:

     node ${CLAUDE_PLUGIN_ROOT}/scripts/evolution-log.js REFLECT "{path} conf {old}→{new} freshness={new_freshness}"
     

   • For state transitions: also log STALE, CONTRADICT events as appropriate.

Never delete or overwrite the belief body or the existing evidence / counter_evidence arrays. Only frontmatter confidence and freshness change.

Dry-run: report intended changes, write nothing.

---

Phase 3 — Promote (deterministic, no LLM)

Skip if --consolidate-only or --reflect-only.

Goal: pin top beliefs into Persona.md ## Top Beliefs based on thresholds. Pure script — no judgement required.

Step 3a: Run the script

If --dry-run:

node ${CLAUDE_PLUGIN_ROOT}/scripts/promote.js --dry-run

Otherwise:

node ${CLAUDE_PLUGIN_ROOT}/scripts/promote.js

The script:

• Filters Notes/*.md with type: belief by configured thresholds (confidence, sources, freshness)
• Ranks by composite score (confidence × log(sources_count + 1))
• Takes top N (default 10)
• Updates Persona.md ## Top Beliefs (wikilinks only, never copies)
• Demotes entries no longer meeting threshold
• Logs PROMOTE / DEMOTE events to ~/.local/state/remember/evolution.log

If auto_promote: false in config, the script is still safe to call — it computes deltas but does not write.

---

Step 4: Report

🧠 Brain evolved (date: {today})

Phase 1 — Consolidate
  Entities re-synthesized: {N}
  - [[People/cezar.md]] (+3 touches)
  - [[Projects/dollie/dollie.md]] (+5 touches)

Phase 2 — Reflect
  Beliefs re-evaluated: {N}
  - {N} stable
  - {N} strengthening
  - {N} weakening (review recommended)
  - {N} stale
  - {N} contradicted

Phase 3 — Promote
  Mode: {bootstrap | normal}  (bootstrap is in effect while total beliefs < 20)
  Effective thresholds: conf>={c} sources>={s}
  Top {N} beliefs pinned to Persona.md.
  Promoted: {N}
  - [[Notes/x.md]] conf=0.91 sources=8
  Demoted: {N}
  - [[Notes/y.md]] (now conf=0.74)

Audit: ~/.local/state/remember/evolution.log

Always show the audit log path so the user can tail it and verify.

---

Error Handling

• Missing brain → tell user to /remember:init
• Missing Persona.md → run promote with auto_promote: false semantics (compute, don't write); warn user
• Empty Notes/ → all phases no-op cleanly
• Invalid frontmatter on a belief → skip that file with a warning, continue with the rest
• Phase 1 partial failure → continue to Phase 2/3 unless aborted explicitly

See @reference.md for detailed phase logic, freshness state machine, and report templates.