| name | capture |
| description | Capture new information into the shared Memory OS. Use when importing a photo, pasted text, artifact, markdown note, research, review, plan, or real-world event and you want memory output, document output, or both. Not for summarizing-only, transcription, or promotion review. |
| argument-hint | [memory|doc|both] [topic, artifact, or pasted content] |
| disable-model-invocation | true |
Capture
Use the shared Memory OS contract at ~/.config/context/AGENTS.md.
Goal
Provide one front door for new information so the user does not need to choose between memory systems.
Default behavior:
- infer whether the input should become
context/, docs/, or both
- ask one short clarifying question only when confidence is low
- preserve provenance when the source is external
Read Order
~/.config/context/AGENTS.md
~/.config/context/docs/memory-os-contract.md
~/.config/context/docs/review-note-contract.md
~/.config/context/docs/repo-profiles.md
~/.config/context/docs/productivity-integration.md
- Relevant repo-local instructions if the target repo is known
Memory Model
Treat these as separate layers with different jobs:
CLAUDE.md -> hot memory for a few broadly relevant repo cues
context/ -> deep recall memory for compact durable context
docs/ -> full authored documents
Do not create or depend on PARA folders as the primary storage model.
Modes
The user can force a mode:
memory -> write only durable memory outputs
doc -> write only authored document outputs
both -> write a full document and a compact memory summary
If no mode is specified, route automatically.
Routing Rules
Prefer context/
Use context/ when the primary outcome is compact recall or decoding support:
- people context
- project summaries
- glossary terms, acronyms, nicknames, aliases
- durable preferences
- critical life or work context extracted from messy real-world inputs
- brief incident summaries where the long-form artifact is not the main thing
Typical homes:
context/glossary.md
context/people/
context/projects/
context/context/
Promote a short pointer into CLAUDE.md only when the information becomes high-frequency and broadly relevant across most sessions in that repo.
People Note Contract
When a capture updates a person, treat context/people/*.md as one canonical note system shared with productivity-sync and /people-enrich.
For people-related captures:
- do not freehand edit
context/people/*.md
- prepare structured JSON and call
~/.claude/skills/people-enrich/scripts/apply-person-update.ts
- write durable observations to
## Signals
- write ambiguity, conflicts, or follow-up review items to
## Open Questions
- keep
## Relationship Profile updates conservative and H3-scoped
- never paste raw message logs or transcript dumps into durable people notes
- prefer a short durable signal over a miniature dossier
Structured payload example:
{
"signals": [
"2026-03-22: Mentioned Easter lunch plan; source: iMessage"
],
"open_questions": [
"Confirm whether the lunch is Easter Sunday or Easter Monday."
],
"source_handles": {
"email": ["mel@example.com"]
}
}
Safe dry-run proof:
bun run ~/.claude/skills/people-enrich/scripts/apply-person-update.ts \
--people-dir ~/code/my-second-brain/context/people \
--source imessage \
--handle +61412667520 \
--report /tmp/person-update.json \
--output /tmp/person-update.preview.md
Review the preview first. Only switch to --write after the proposed note matches the contract.
Prefer docs/
Use docs/ when the primary outcome is an authored document the user will want to read, edit, or revisit in full:
- research notes
- review notes
- plans and itineraries
- specs
- decisions
- logs
- artifact transcriptions or structured writeups
Typical homes:
docs/research/
docs/reviews/
docs/plans/
docs/specs/
docs/decisions/
docs/logs/
docs/artifacts/
Use these nearby note families deliberately:
review for evaluating a change, artifact, system, or workflow against explicit criteria
plan for evolving options, sequencing, itineraries, and future-facing decision documents
spec for more settled canonical documents that describe the shape to execute against
Use both
Use both outputs when the material has a long-form document and a compact durable summary:
- trip plans with an ongoing summary note
- medical or family events with a full incident note plus key durable context
- research that should also update a project or person memory note
- imported markdown files that should be preserved as docs but surfaced as concise memory
In both mode:
- create or update the full document in
docs/
- create or update a compact summary in
context/
- update
CLAUDE.md only if the information is high-frequency enough to matter in daily decoding
Repo Resolution
Before writing anything:
- Resolve the owning repo.
- Keep repo-specific memory local by default.
- Use
my-second-brain only for life-hub material, cross-project synthesis, or personally durable memory.
- If the current repo is not clearly the owner, do not assume it is.
Classification Questions
For every capture, decide:
- Who owns this?
- Is the primary output compact memory, an authored document, or both?
- Is the material semantic, episodic, or runtime?
- What is the smallest useful note family?
- What provenance should be preserved?
- Should any part of this be promoted into
CLAUDE.md as a short hot-memory pointer?
Note Families
Map captures back to the shared note families whenever possible:
project
area
person
pet
meeting
research
review
task
decision
adr
plan
spec
runbook
log
artifact-sidecar
Use status to represent lifecycle such as planned, active, done, or archived.
Minimum Frontmatter
Use the smallest shape that helps retrieval:
---
title: "Human-readable title"
type: note-type
status: active
updated: YYYY-MM-DD
summary: "Optional retrieval summary"
source: /path/to/source
source_system: photos|email|calendar|drive|repo
source_id: external-id
source_url: https://example.com
related:
- path/to/related-note.md
---
Provenance Rules
Preserve provenance whenever the material comes from:
- pasted screenshots or photos
- email
- calendar
- trackers
- drive documents
- imported markdown files
- copied web content
Prefer the smallest useful provenance fields. Do not scatter raw links across many notes.
If an imported file becomes the new canonical copy in the owning repo, do not keep a stale source: pointer to a retired repo just because the content originated there. Keep provenance only when it still helps future retrieval or auditing.
Workflow
Step 1: Parse Intent
Check whether the user explicitly asked for:
If not explicit, infer the best route from the input.
Step 2: Resolve Owner
Identify the source-of-truth repo before writing anything.
Step 3: Choose Output Shape
Pick one:
Step 4: Place The Result
Route by content type:
- people, glossary, project summaries, preferences ->
context/
- research, reviews, plans, specs, decisions, logs, artifacts ->
docs/
- mixed cases -> both
When writing context/projects/, use lowercase hyphenated slugs and prefer domain-prefixed names when they improve findability, such as holiday-fnq-july-2026.md, not generated names like project_fnq_july_2026.md.
Step 5: Preserve Provenance
Add source fields or create an artifact-sidecar when needed.
Step 6: Consider Promotion
Only add or update CLAUDE.md when the captured information will materially help day-to-day decoding or execution.
Default Heuristics
Use these defaults unless the user says otherwise:
- photo of a referral, form, receipt, or message ->
memory or both
- pasted shorthand, acronym, nickname, person context ->
memory
- research dump ->
doc
- imported markdown plan or itinerary ->
both
- meeting artifact that needs a proper writeup ->
doc or both
- durable project summary update ->
memory
- imported planning documents that remain option-heavy ->
doc or both with type: plan
- imported planning documents that are now the settled canonical operating copy ->
doc or both with type: spec
Shape Heuristics
Use this decision ladder:
- Inline
- Use when the content is short, stable, and best understood inside the parent note.
- Good default for simple bookings, one-off references, and compact status detail.
artifact-sidecar
- Use when the content is mainly an external artifact with provenance.
- Good for confirmations, receipts, PDFs, email threads, transcripts, and exported reports.
- Strong signal: IDs, dates, terms, links, attachments, or status detail worth preserving.
- Standalone note
- Use when the subtopic has independent meaning or follow-up.
- Strong signal: more than one expected update, more than one artifact, its own tasks or decisions, or likely direct search/linking later.
When uncertain, choose the smaller shape and mention when to split it later.
Examples
Example: Real-World Artifact
Input:
- photo of a doctor referral for Levi after an emergency visit
Route:
- default to
both if the event matters beyond the raw image
Write:
- durable family/medical context in
context/
- dated incident or artifact note in
docs/logs/ or docs/artifacts/
Example: Research Session
Input:
- pasted research notes about a product or workflow
Route:
Write:
docs/research/YYYY-MM-DD-topic.md
- optional
context/ summary only if it becomes durable and cross-context
Example: Travel Plan Import
Input:
- a batch of markdown itinerary files
Route:
Write:
- full itinerary docs in
docs/plans/
- compact trip summary in
context/projects/
Output Shape
When asked to recommend placement, respond with:
- Recommended home
- Output mode:
memory, doc, or both
- Memory layer: semantic, episodic, or runtime
- Note family
- Proposed paths
- Minimum frontmatter
- Provenance needs
CLAUDE.md promotion recommendation
Rules
- Prefer one capture flow over multiple user-facing routing decisions.
- Prefer source-of-truth repos over dumping into
my-second-brain.
- Prefer the smallest useful metadata shape.
- Prefer
docs/ for long-form authored material.
- Prefer
context/ for compact durable recall.
- Use
both when the user benefits from a full document plus a concise memory summary.