obsidian-notes

name	obsidian-notes
description	Use when taking notes, recalling context, building connections, or treating the Obsidian vault as Claude's persistent memory across sessions

Obsidian Notes — Claude's Persistent Memory

When to Activate

• User asks to remember, note, or document something
• A non-obvious solution is found worth preserving
• Cross-domain connections surface
• Starting work on a topic that may have prior vault context
• End of a meaningful work session (session log)
• Before /compact (pre-compact checkpoint)
• User mentions the vault, notes, or Obsidian

---

Philosophy

The vault is Claude's external brain — persistent, cross-session memory where thinking accumulates. Write with intent, not volume: direction/decision notes only on user-approved direction, implementation notes on a need-to-know basis, plumbing freely (see the Write Policy in the vault rules). Obsidian is the overflow for insight that would otherwise vanish at session end.

---

Vault Structure

~/obsidian_notes/
├── inbox/       Quick capture. Process later. Default drop zone.
├── projects/    Time-bound work (grouped by project: frontmatter, no hub files).
│   └── <project>/   ← subfolder per project when 2+ notes exist
├── areas/       Durable domain knowledge (promoted from projects/ when reused).
├── library/     Atomic reference notes, papers, tools.
├── personal/    Goals, journal, personal context.
└── agent/       Agent's own synthesis, connections, open questions.
    ├── session-log.md      Running log (append-only)
    ├── open-questions.md   Unresolved questions (append-only)
    ├── connections.md      Cross-domain links (append-only)
    └── instincts.yaml      Learned behavioral patterns with confidence

Subfolder Rules

• 1 note → place directly in parent folder (projects/some-topic.md)
• 2+ notes → create subfolder (projects/project/architecture.md)
• No _index.md hub files. Project grouping uses project: frontmatter. Agents find notes via vault_project / vault_find (Grep on project: <name> is the last resort).
• Do NOT create generic hub filenames (Home.md, _index.md, index.md). Use descriptive project-prefixed names if a hub is needed (e.g., ece8893-fpga-overview.md).
• Never prefix-based flat naming (projects/proj-arch.md, projects/proj-bugs.md)
• Folder depth: nest as deep as needed — the vault tools handle arbitrary depth — but prefer 1–2 levels (e.g. projects/project/note.md) for readability
• Area promotion: When a second project needs knowledge from the first, extract the shared concept to areas/.

---

Quality Gate (MANDATORY before writing any note)

Before creating or modifying a vault note, check ALL of these:

• Write class: Direction/decision note? → needs explicit user approval first. Implementation note? → need-to-know only, isolated from direction notes. Plumbing/inbox? → autonomous.
• Future value: Would a future Claude instance genuinely benefit?
• No duplication: Not already in codebase, git history, or existing notes?
• Frontmatter: Has date, tags, type, and status in YAML frontmatter?
• Wikilinks: Contains at least one [[wikilink]] to related note?
• Correct folder: projects/ for time-bound work, areas/ for durable knowledge, library/ for reference?
• Filename: Lowercase-hyphenated, no spaces/CamelCase?
• Size: Under 800 lines? (Split if larger, target 200-400)
• Content: Captures reasoning/tradeoffs, not just facts?

---

Note Format

Required Frontmatter

---
date: 2026-04-01
tags: [ece, synthesis]
type: concept          # concept | decision | log | mission
status: active         # backlog | active | completed | archived
project: my-project    # optional, groups notes without hub files
---

Every note except running logs (session-log.md, open-questions.md, connections.md) MUST have frontmatter.

Naming

• YYYY-MM-DD-topic.md — time-anchored notes
• topic.md or topic-subtopic.md — evergreen reference
• Lowercase-hyphenated only. No spaces, CamelCase, or uppercase.

Content

• Use [[wikilinks]] to connect related notes
• Use headers to structure longer notes
• Prefer atomic, linkable notes over monolithic documents
• Capture reasoning and tradeoffs, not just conclusions

Appending vs Overwriting

• Prefer appending — context accumulates over time
• Use ## YYYY-MM-DD headers for timestamped additions
• Only overwrite when restructuring with user awareness

---

Context Modes

Adapt behavior based on what the user needs:

Capture Mode (default)

Quick notes, minimal processing, inbox-first. Use when the user shares something worth remembering but isn't doing deep work.

Synthesis Mode

Deep reading, cross-note connection building, gap identification. Activate when the user asks to "review", "connect", or "synthesize" vault content.

Review Mode

Vault audit / quality checks → load the obsidian-audit skill instead.

---

Session Log Pattern

At the end of meaningful sessions, append to agent/session-log.md:

## YYYY-MM-DD
**Worked on:** [brief description]
**What worked:** [approaches that succeeded, with evidence]
**What failed:** [approaches that didn't work, and why]
**Key decisions:** [decisions made and rationale]
**Open:** [unresolved items]
**Connections:** [[note1]] ← [[note2]] [brief link description]

Skip the log if the session was trivial (quick Q&A, simple edit).

---

Connection-Building

When writing a note, always ask: "What other notes does this connect to?"

connections.md Format (with confidence)

## YYYY-MM-DD
**Connection:** [[note1]] ↔ [[note2]]
**Insight:** [the cross-domain link — what makes this non-obvious]
**Confidence:** high/medium/low
**Evidence:** [what supports this]

Only add genuine cross-domain connections. Forced connections degrade signal.

---

Instincts (Learned Patterns)

agent/instincts.yaml tracks behavioral patterns learned across sessions. Each instinct has a confidence score (0.0-1.0) that evolves:

• Increases when the pattern is validated in practice
• Decreases when contradicted
• Pruned below 0.3 confidence

When you notice a reusable pattern, add it. When an instinct proves wrong, lower its confidence or remove it.

---

Pre-Compact Checkpoint

Before any /compact: ensure in-progress notes are written; capture current task state, key paths, and next steps in the session log or a quick inbox/ note; use a descriptive compact summary (/compact Focus on implementing X next).

Survives: CLAUDE.md, tasks, memory files, git state, disk files. Lost: intermediate reasoning, file contents read, conversation history.

---

Retrieval Pattern

Use the native vault MCP tools as your primary retrieval interface — they read a cached graph index instead of grepping the vault.

1. First pass: vault_project(name=<name>) for a project map; vault_find(query="<keyword>") for keyword/name lookup; vault_semantic_search(query="<concept>") to retrieve the exact paragraphs discussing a concept (cross-project, with file + line range).
2. Evaluate: Scan results (title + frontmatter + link counts + snippet). Pick 1-3 candidates.
3. Follow links: vault_links(note=<note>) to find adjacent context.
4. Read content: Only Read the 1-3 notes that look most relevant.

Fall back to Grep only for full-text patterns inside note bodies. Also check agent/session-log.md (past work) and agent/open-questions.md (unresolved items).

---

Sync

A 5-minute scheduler auto-commits — no manual push needed. Force immediate sync: ~/.agent-configs/hooks/server-sync.sh.