remember

Remember — Brain Dump Skill

Immediate capture: when the user says "remember this", "save this", "brain dump", etc., route content to the right place in the Second Brain.

⚠️ Built-in Tools Only (NO Bash for file ops!)

Operation	✅ Use This	❌ NOT This
List files	LS tool	bash ls
Find files	Glob tool	bash find
Search content	Grep tool	bash grep
Read files	Read tool	bash cat
Create files	Write tool	bash echo >
Update files	Edit tool (old_string → new_string)	bash sed / rewrite

Only use bash for: running build-index.js.

---

Brain Dump Pipeline

Step 1: Get Knowledge Index

node ${CLAUDE_PLUGIN_ROOT}/scripts/build-index.js --compact

Use this to prevent duplicates and enable smart linking.

Step 1b: Check REMEMBER.md

Read REMEMBER.md files (cascading):

1. {brain}/REMEMBER.md (global preferences)
2. ./REMEMBER.md in current project (project-specific, if exists)

Apply all rules from both. Project additions layer on top of global. User instructions override defaults.

Step 2: Parse User Input

Extract from conversational input: person names, project references, tasks, decisions, learnings, area updates, URLs.

Step 2.4: Journal the capture FIRST

Before creating or editing any L2 file, append the user's verbatim quote to {brain}/Journal/{TODAY}.md (create the file if it doesn't exist; use the daily template from reference.md). This becomes the source-of-truth for the evidence: field on every L2 file you create in this capture.

Format inside the journal file:

## {Topic / Project name}
- "{verbatim user quote}"

The journal entry's filesystem path is what every evidence.source will reference. Never invent a chat/... or session/... path that doesn't correspond to a real file.

Step 2.5: Tag with epistemic type

For every chunk of content extracted in Step 2, classify it as one of:

• world-fact — verifiable claim about reality (decision, technical fact, learning)
• belief — subjective claim with implicit confidence (preference, hypothesis, judgment)
• observation — synthesized profile per entity (creates/updates People/<x>.md, Projects/<x>/<x>.md, Areas/<x>.md)
• experience — first-person event log (always Journal/.md)

Heuristics (apply in order; first match wins):

1. Contains a date or "met with" / "called" → experience
2. Phrase patterns "we decided", "going with", "chose X over Y" → world-fact
3. Predicate about an entity ("works as", "lives in", "based in", "leads the team") → observation
4. Words "prefers", "probably", "I think", "seems like", "better to" → belief
5. Else default → world-fact

When creating or editing a file, write/preserve the frontmatter type: field accordingly. For beliefs, confidence is REQUIRED (use your best estimate 0.0–1.0).

The same rules live in scripts/schema.js (detectType); use that as canonical reference if uncertain.

Step 3: Build Resolution Map

Resolve every name/reference against the knowledge index:

• Matches existing → Edit tool to update
• New entity → Write tool to create
• Ambiguous → ask user

Fuzzy matching: "John", "john smith", "John S." → People/john-smith.md if exists.

Step 3.5: Dedup check (before any new belief/world-fact write)

Before creating a new Notes/<slug>.md for a belief or world-fact, check if a similar one already exists:

node ${CLAUDE_PLUGIN_ROOT}/scripts/append-evidence.js find-similar {brain} <slug> belief

Or for world-fact: pass world-fact as the type filter.

If no similar file → proceed to Step 4 (Write new).

If a similar file is returned → run the polarity check (Step 3.6) BEFORE appending.

Step 3.6: Polarity check (only when Step 3.5 found a similar belief)

Read the existing file (its body + the evidence: array). Compare the new claim's stance to the existing belief's stance. There are exactly three outcomes:

1. Same direction (re-affirms existing belief) → append as positive evidence:

   node ${CLAUDE_PLUGIN_ROOT}/scripts/append-evidence.js append <filepath> '{"source":"Journal/<TODAY>.md","quote":"<verbatim>","date":"<TODAY>"}'
   

   This increments sources_count, appends to evidence:, updates updated:, refuses duplicate sources (idempotent).

2. Opposite direction (contradicts existing belief) → append as counter-evidence and log the event:

   node ${CLAUDE_PLUGIN_ROOT}/scripts/append-evidence.js append-counter <filepath> '{"source":"Journal/<TODAY>.md","quote":"<verbatim>","date":"<TODAY>"}'
   node ${CLAUDE_PLUGIN_ROOT}/scripts/evolution-log.js CONTRADICT "<filepath> counter from Journal/<TODAY>.md"
   

   The first command appends to counter_evidence:, leaves sources_count untouched, and automatically flips freshness: contradicted when counter entries outnumber positive ones. The second writes a CONTRADICT line to ~/.local/state/remember/evolution.log for audit.

3. Unrelated (different topic, fuzzy match was wrong) → ignore the find-similar match and create a new file via Step 4.

Polarity rules — apply in order, first match wins

• Explicit negation of the existing claim (e.g. existing: "prefer Postgres for small projects", new: "actually SQLite is better for small projects") → counter.
• Reversal of preference (existing: "X over Y", new: "Y over X") → counter.
• Same subject, opposite verdict (existing: "Redis is overkill for X", new: "Redis is the right call for X") → counter.
• Restatement, paraphrase, or new evidence in the same direction → positive.
• Adjacent topic, same domain, neither confirms nor contradicts → ignore find-similar; create a new note.
• You are not at least 80% confident the polarity is opposite → default to positive evidence. Don't guess.

Worked examples

• Existing Notes/pref-postgres.md (belief, "prefer Postgres for small projects, avoid Redis").

  • New: "going with Postgres for the dollie project" → positive (same direction, new instance).
  • New: "actually SQLite is better than Postgres for small projects" → counter (explicit reversal).
  • New: "Redis works great for the dollie cache layer" → ignore find-similar (different subject — Redis-as-cache vs Postgres-as-DB are not the same belief).

• Existing Notes/pref-async-comms.md (belief, "prefer async comms over meetings").

  • New: "meetings are essential for kickoffs" → counter only if existing belief is unconditional; if the existing belief already excludes kickoffs, treat as ignore.
  • New: "Slack > meetings for status updates" → positive (paraphrase).

If polarity is genuinely ambiguous, default to positive evidence and add a brief note in the chat confirmation so the user can re-route manually. Never guess "counter" — false contradictions corrupt the brain faster than missed ones.

Step 4: Route & Write

For existing files → use Edit tool (surgical updates, not rewrites). For new files → use Write tool with templates from reference.md.

Check REMEMBER.md ## Templates section for overrides before using defaults.

Use [[wikilinks]] everywhere. Link forward; Obsidian handles backlinks.

See reference.md for detailed templates and routing tables.

Step 5: Auto-promote (deterministic, every capture)

After all writes are done, run promote.js once. It is deterministic, fast, and zero LLM cost — so it can run on every capture without any user opt-in. This keeps Persona.md ## Top Beliefs in sync with the brain in real time.

node ${CLAUDE_PLUGIN_ROOT}/scripts/promote.js

The script no-ops if nothing crossed the threshold. If something promoted or demoted, surface the delta in Step 6.

Step 6: Confirm

✅ Brain updated:
  - Updated People/john-smith.md (+interaction)
  - Created Notes/decision-nextjs.md
  - Updated Tasks/tasks.md (+1 Focus)
  - Updated Journal/2026-02-15.md (+1 section)
  - Notes/pref-postgres.md (+counter_evidence — freshness flipped to contradicted)
  - Persona.md ## Top Beliefs: +1 promoted ([[Notes/pref-async-comms.md]])

If a polarity decision was non-obvious (close call between positive and counter), say so in the confirmation so the user can correct it before the next capture.

If the auto-promote step skipped writing (e.g. auto_promote: false in user's config, or no Persona.md), don't bother mentioning it — silent no-op is the right UX.

---

Content Routing

Content	Destination
Person interaction	People/{name}.md
Task with deadline	Tasks/tasks.md (Focus)
Task without deadline	Tasks/tasks.md (Next Up)
Future/roadmap task	Projects/{name}/{name}.md
Project update	Projects/{name}/{name}.md
Decision	Notes/decision-{topic}.md
Learning/insight	Notes/{topic}.md
Area update	Areas/{area}.md
URL/resource	Resources/{type}/{title}.md
Unclear	Inbox/{topic}.md

File Naming

• Folders: kebab-case/
• Files: kebab-case.md
• People: firstname.md or firstname-lastname.md
• Dates: YYYY-MM-DD.md

Schema rules

• Every L2 file (Notes/People/Projects/Areas) carries type:. Use Step 2.5 heuristics or scripts/schema.js detectType().
• Every captured fact must include at least one evidence entry: { source: <real-file-path>, quote: <verbatim>, date: <SESSION_DATE> }.
• source: MUST be a real file path inside the brain — typically Journal/<SESSION_DATE>.md (the journaled capture from Step 2.4). Never invent paths like chat/..., session/..., or anything that doesn't exist on disk.
• For type: belief, confidence: 0.0–1.0 is REQUIRED.
• freshness: stable is the default for new captures. The evolve skill (Phase 2) updates this later.
• Never overwrite L2 files. New positive evidence → evidence:. New contradicting claim → counter_evidence:. Polarity is decided live in Step 3.6 — don't defer to /remember:evolve.

Validate after write

After every Write or Edit on a brain file (Notes/, People/, Projects/, Areas/, Journal/, or Persona.md), run the validator:

node ${CLAUDE_PLUGIN_ROOT}/scripts/schema.js validate <filepath>

The output is JSON: {changed, addedFields, addedSections, warnings}.

• If changed: true, the file was upgraded with missing schema fields/sections — note this in your final report.
• If warnings contains items (e.g. "confidence defaulted to 0.5 — review"), surface them in your final response so the user can refine the value.
• If changed: false and no warnings, your write was already complete — nothing else to do.

Aim to emit complete frontmatter on first write so the validator is a no-op. Skip validation for files outside the brain (Inbox/, Tasks/, Archive/, Templates/) — the validator returns passthrough anyway, but you can save the call.