brain-ingest

name	brain-ingest
description	Import external sources into the brain — articles, videos, papers, web pages, handover documents. Use when the user asks to ingest, import, or add external content to the knowledge base, or run /brain-ingest.

Brain Ingest

Import external sources into the brain — articles, videos, papers, web pages. This is Karpathy's "raw → wiki" ingest operation.

Trigger

Activate when the user asks to:

• Ingest, import, or add an external source to the brain/knowledge base
• Save an article, video transcript, paper, or web page for future reference
• Run /brain-ingest

Workflow

1. Capture the source

Determine the source type and obtain the content:

• URL: Fetch the page content using WebFetch or browser tools
• YouTube video: Use the youtube-transcript skill to get the transcript
• File path: Read the file directly
• Pasted text: Use the content as-is

2. Save raw source to inbox

Write the raw content to ~/.brain/capture/inbox/ as a provenance-tagged capture. Filename: YYYY-MM-DD-<source-slug>.md. Include frontmatter:

---
type: capture
title: "<source title>"
sources:
  - <original URL or path>
created: <today>
agent_last_edit: <agent>
---

For content mirrored from an external system (e.g. via the service's mirror agent), the frontmatter also carries provenance fields written by the mirror pipeline:

---
type: capture
title: "<entry title>"
source: "<target-name> (<target-kind>)"
mirrored_at: "<ISO-8601 timestamp>"
mode: "full" | "index-only"
pointer: "<target-name>:<relpath-or-id>"
---

Do not write these provenance fields manually when ingesting from a URL or file — they are reserved for the service mirror path.

3. Trigger live processing when enabled

After writing the inbox capture, call the configured Fritz Local processing path:

• If settings.local_brain_service.enabled: true and auto_compile_on_ingest is not false, trigger the Local Brain compile path (service first, local fallback) so the capture does not silently pile up.
• If processing is not active, leave the raw capture in ~/.brain/capture/inbox/ and report that compile is pending/manual.

If live processing is unavailable and manual compile is needed, read the raw source and create a compiled knowledge article. In registry-free mode, write to ~/.brain/knowledge/<scope>/<section>/<slug>.md (scope = common or project slug; section = decisions, lessons, runbooks, or context). In vault mode, write to the appropriate knowledge/ subfolder. Either way:

• Extract key concepts, facts, and insights
• Structure as a clear, scannable article
• Cross-reference with existing articles where relevant
• Include full frontmatter with sources pointing to the inbox file

4. Update index and log

• Add the new article to the index file
• Append to .brain/log.md:

YYYY-MM-DD HH:MM | INGEST | <agent> | Ingested "<title>" from <source type>, created <article path>

Example

User: "Ingest this video: https://youtu.be/1FiER-40zng"

1. Fetch transcript via youtube-transcript skill
2. Save raw transcript to capture/inbox/2026-04-07-cole-medin-second-brain.md
3. Extract key concepts: memory layer, hooks, skills, heartbeat, zero-trust security
4. Create knowledge/300_Resources/ai-second-brain-architecture.md with structured summary
5. Update index, log the operation

Important

• Always preserve the raw source in capture/inbox/ — it's the provenance chain
• The knowledge article is a distillation, not a copy — extract insight, don't dump text
• Cross-reference existing articles using related frontmatter