Brain Capture

Commit something to memory. I want to remember this.

If the user provided text: $ARGUMENTS

If $ARGUMENTS is empty, ask what they want me to remember. Help them structure it:
- **Decision:** "Decided to [X] because [Y]. Alternatives considered: [A, B]."
- **Person note:** "[Name] — [role/context]. Talked about [topic]. Key takeaway: [insight]."
- **Meeting:** "Met with [people] re: [topic]. Decided: [list]. Actions: [list]."
- **Preference:** "Always [do X] / Never [do Y]. Reason: [why]."
- **Pattern:** "[Approach] works well for [situation]. Learned this when [context]."
- **Impression:** "[Person/system] is [observation]. Evidence: [what I noticed]."

Capture by running open_brain.py with the --capture flag. Use the installed path (`~/.claude/hooks/open_brain.py`) or the repo path (`scripts/open_brain.py`) — whichever exists.

```bash
python3 ~/.claude/hooks/open_brain.py --capture "<formatted thought>" --source "claude-code" --session-id "$CLAUDE_CODE_SESSION_ID" --project "<current_project_name>"

Replace <current_project_name> with the actual current project/directory name.

After capturing, briefly confirm what I remembered and what metadata was extracted.

Seeding NAL truth values (T2.6)

By default, the truth-value {frequency, confidence} (stv) is derived from the confidence label Haiku extracts from the capture text (high→c=0.9, medium→c=0.7, low/absent→c=0.5; frequency defaults to 1.0). You can override this explicitly:

• --stv-f FREQ — set stv frequency (0.0–1.0). Use 0.0 for a fully-refuted belief, 1.0 for strong positive evidence, values between for partial support.
• --stv-c CONF — set stv confidence (0.0–1.0). Represents weight of evidence; higher = more observations backing this belief. Confidence of 0.35 or below causes search results to display a [LOW-CONFIDENCE] marker.

python3 ~/.claude/hooks/open_brain.py \
  --capture "Postgres HNSW index recall ≥ 0.95 at 1M vectors (measured 2026-05-01)" \
  --stv-f 1.0 --stv-c 0.85 \
  --source "claude-code" --session-id "$CLAUDE_CODE_SESSION_ID" --project "optivai-builder"

These flags pair with /brain-revise: if you later find a contradicting atom, --revise will fuse the two stv values via NAL evidential-horizon revision. Seeding accurate stv-c values at capture time makes future revisions more meaningful — the weighted average is only as good as the input confidence values.

Optional PROV-DM fields (v0.2.0+)

For explicit W3C PROV-DM provenance, the following flags are supported. The default capture path stamps these automatically from session context, but you can override when the thought derives from a specific upstream source (e.g. summarising a meeting note, transforming a prior decision):

• --prov-agent <name> — who/what produced this thought (e.g. claude-code, ralph-loop, chris-manual).
• --prov-activity <verb> — the producing activity (capture, summary, synthesize, transform).
• --derived-from <thought_id> — the parent thought this one derives from; populates was_derived_from and makes the new thought walkable by /brain-trace.

Use these when capturing a derived thought you expect to surface in a citation chain. /brain-trace will then walk the --derived-from pointer back to the source.

Why this command exists in the neurosymbolic discipline

This is the agent's instrument for Rule 3 — Capture-with-alternatives and the registration half of Rule 5 — Skill-lifecycle. The MS_ε primitive enacted is WA (Write Authorization): every capture passes schema validation, RE2 PII redaction, and PROV-DM stamping before the atom commits. Anonymous writes are rejected at the type level — the resulting atom carries {agent, activity, wasGeneratedBy, wasDerivedFrom?, sourceUri?} so /brain-trace can walk its lineage later. Capture is the move that turns reasoning into institutional memory; skipping it discards the audit trail every prior agent has been building.

When to invoke

• After making a decision of consequence — architectural choice, dependency selection, an approach that closes off other paths. The capture body MUST include alternatives_rejected: with the choices not taken (Rule 3). Plain capture without alternatives reads as "I picked X" and surfaces no learning to future agents.
• When you discover a reusable pattern — a sequence of moves that worked for a class of problems and is likely to apply again. Capture as type=pattern, then call /brain-promote <id> so the next similar task recalls it before re-deriving (Rule 5).
• When the user states a preference, correction, or "always do X / never do Y" rule — capture as type=preference immediately so the next session inherits it.
• When you learn something about a person — capture as type=person_note with the context, role, and the specific signal you observed.
• When two recalled atoms conflict and you resolve them via NAL revision, capture the fused belief with --derived-from pointing at both premises so the resolution is auditable (Rule 2).

How to use the result

• Confirm the returned thought_id and the metadata Haiku extracted (type, topics, people, summary, confidence). If extraction got the type wrong, immediately re-capture with the corrected framing rather than letting the wrong kind drift downstream.
• For derived captures, verify the was_derived_from field is populated by running /brain-trace <new_id> and confirming the chain walks back to the intended parent.
• If the captured atom matters more than its raw similarity score will likely surface, call /brain-promote <id> once after capture to seed Hebbian weight (Rule 4).
• If the user immediately corrects the capture ("no, the alternative I rejected was different"), do NOT call /brain-forget; capture a corrective atom with --derived-from pointing at the wrong one and /brain-demote the wrong one. Forgetting is reserved for the user.

Pearl atom-kind hint

Capture kind selection ```