Claude's long-term knowledge graph, your DEFAULT memory destination. Invoke this skill before any brain read, write, search, or update. TRIGGER when: user mentions "brain", "engram", or "knowledge graph"; asks to remember, recall, or check something; you learn something with any depth or connections; wrapping up a session with notable outcomes; starting a session where prior context would help; you encounter or reference an external source worth preserving; a discussion produces insights worth grounding; user shares a link in conversation; a tier-2 digest (in built-in memory) points into engram via `engram:` frontmatter or a `[[wikilink]]` and you need the full depth, backlinks, or related decisions. Do NOT trigger for: quick one-off questions, tasks that don't produce reusable knowledge, or when the user says to skip memory/brain.
Your long-term knowledge graph lives at ~/engram/ as an Obsidian vault. This is your default memory destination, not opt-in, not "nice to have." If information has depth, connections, or reasoning behind it, it goes here.
All non-Sources vault access goes through obsidian:obsidian-cli, always use vault="engram" and silent when creating notes. Scripts handle Sources/ (see below).
Helper scripts are in the scripts/ subdirectory next to this skill file. Find this skill's directory first, then reference scripts from there:
<skill-dir>/scripts/archive-source.sh: download and archive URLs/files to Sources/<skill-dir>/scripts/crawl-thread.py: crawl X/Twitter threads<skill-dir>/scripts/clean-transcript.py: clean VTT subtitles<skill-dir>/scripts/lint-brain.sh: vault health scan| Tier | System | Purpose |
|---|
| Depth |
|---|
| Always loaded? |
|---|
| 1. Index | MEMORY.md | One-line pointer per notable note + flat orientation facts | Minimal | Yes |
| 2. Brief overview | Built-in memory sibling files (project_*.md, user_*.md, feedback_*.md, reference_*.md) | 10–30 line digest: what it is, why, current status, key tech, pointer to engram for depth | Medium | No, loaded via Read when MEMORY.md entry is followed |
| 3. Knowledge graph | Obsidian vault (~/engram/) | Full architecture, decisions, session history, sources, and the [[wikilink]] connections between them | Full | No, queried via obsidian-cli when backlinks or deeper context needed |
Each tier holds ~10× more detail than the previous. The three-tier shape means context is always available at the appropriate resolution, a one-liner in every session (free, via always-loaded MEMORY.md), an orientation digest after one cheap file read, and the full graph whenever the conversation touches the project at all.
Critical: tier 2 is an orientation primer, not an answer source. Digests are frozen at creation time; anything about current status, recent work, decisions, or connections lives in engram. If the task is actually about the project (status, architecture, updates, edits), pull engram, tier 2 alone will lead you to hedged, partial answers.
Routing heuristic:
project_*.md holds a brief digest that points into engram; MEMORY.md holds the one-liner.The critical rule: tier 2 is a digest, not a shadow copy. It summarizes; engram contains. When depth is needed, Claude follows the digest's pointer into engram, it does not duplicate what's in engram at tier 2.
Don't ask "should I save this?", just save it to the right tier(s).
MEMORY.md is loaded into Claude's context on every turn. That means tier 1 (the index) and, via Read, tier 2 (brief digests) are both cheap to access, while tier 3 (engram) is reserved for when depth is actually needed. The three tiers work together:
Tier 1, MEMORY.md index format:
## Projects
- [Lens](project_lens.md): lens.jaedynchilton.com: NASA photo analytics, sql.js + FTS5 (2026-04-12)
- [VAI Assistant](project_vai.md): Local agentic AI for Ignition SCADA (active 2026-04-07)
## User
- [Technical breadth](user_interests_and_breadth.md): AI engineering, photography, local ML
- [Tools & environment](user_tools_and_environment.md): M1 Pro 32GB, Docker, Obsidian, tmux
## Feedback
- [Skip planning when urgent](feedback_skip_planning_when_urgent.md): Don't brainstorm when clearly scoped
## References (external systems)
- [Linear INGEST project](reference_linear_ingest.md): pipeline bug tracker
Entries always point to the tier 2 sibling file. Claude can follow the link with Read to get the digest, which in turn points into engram. System prompt truncates after ~200 lines, keep the index tight.
Tier 2, built-in memory sibling file (e.g. project_lens.md):
---