菜单
Save and retrieve agent memory in the Obsidian vault using agent-notes CLI. Defines the single record format for all memory notes. Use when saving decisions, patterns, or session state to the vault.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Review the changes since a fixed point (commit, branch, tag, or merge-base) along two axes — Standards (does the code follow this repo's documented coding standards?) and Spec (does the code match what the originating issue/PRD asked for?). Runs both reviews in parallel sub-agents and reports them side by side. Use when the user wants to review a branch, a PR, work-in-progress changes, or asks to "review since X".
Browser-test handoff via a profile-aware file bus between this dev session and a `claude --chrome` session. Use after frontend work is locally testable or on /chrome-test.
Shared vocabulary for designing deep modules. Use when the user wants to design or improve a module's interface, find deepening opportunities, decide where a seam goes, make code more testable or AI-navigable, or when another skill needs the deep-module vocabulary.
Diagnosis loop for hard bugs and performance regressions. Use when the user says "diagnose"/"debug this", or reports something broken/throwing/failing/slow.
Build and sharpen a project's domain model. Use when the user wants to pin down domain terminology or a ubiquitous language, record an architectural decision, or when another skill needs to maintain the domain model.
A relentless interview to sharpen a plan or design.
| name | obsidian-memory |
| description | Save and retrieve agent memory in the Obsidian vault using agent-notes CLI. Defines the single record format for all memory notes. Use when saving decisions, patterns, or session state to the vault. |
| group | process |
| requires_memory | obsidian,wiki |
Use when you need to persist a discovery, decision, or pattern for future sessions, or retrieve existing knowledge from the vault.
This skill is the single source of truth for memory record format. The CLI implements these rules; do not deviate.
All notes use YYYY-MM-DD_<slug>.md. The folder encodes the type — no type segment in the filename.
type=session): YYYY-MM-DD_<session-id>.md
~/.claude/projects/<cwd-slug>/<session-id>.jsonlidYYYY-MM-DD_<slug>.md (slug of the title)YYYY-MM-DD_<slug>.md where slug is the slugified titleCollision rule: if the target filename already exists, _HHMMSS is appended before .md (e.g. 2026-04-30_fix-foo_142231.md).
A session note is per-session, not per-message. Re-running memory add with type=session for the same session appends an ## Update <UTC ISO> block to the existing file rather than creating a new one.
Every record uses the same template. No date: field — only created_at:.
---
created_at: 2026-04-28T19:30:35Z # ISO 8601, UTC, "Z" suffix — REQUIRED
type: pattern # pattern|decision|mistake|context|session — REQUIRED
session: 2026-04-28_<session-id> # wikilink target (no brackets); absent on session notes themselves
agent: lead # optional
---
After the frontmatter, every note body is followed by:
## Related
(empty initially; Obsidian links can be added manually)
Session notes additionally grow a ## Linked notes section that the CLI populates automatically when non-session notes are written during the same session.
All timestamps everywhere are UTC. Local time is never written. The CLI uses datetime.now(timezone.utc).
Always use the CLI — never write vault files directly:
agent-notes memory add "<title>" "<body>" [type] [agent]
Types:
pattern — reusable solution or technique discovered in the codebasedecision — architectural choice with rationalemistake — recurring error to avoidcontext — project background, constraints, stakeholder notessession — current session's running log (one file per session, appended on each call)Agent: your agent name (lead, coder, reviewer, etc.)
Examples:
agent-notes memory add "Rails enum prefix" \
"Always use _prefix: true to avoid method name collisions" \
pattern coder
agent-notes memory add "Auth middleware rewrite rationale" \
"Driven by legal/compliance requirements around session token storage, not tech debt" \
decision lead
Save when you discover something non-obvious that would cost future sessions time to re-derive:
Do NOT save:
git log, grep)agent-notes memory list # all notes by category
agent-notes memory show <agent> # one agent's notes
agent-notes memory vault # confirm storage and path
The vault is structured as:
vault/agent-notes/projects/
├── <project-name>/
│ ├── Patterns/ — reusable solutions YYYY-MM-DD_<slug>.md
│ ├── Decisions/ — architectural choices YYYY-MM-DD_<slug>.md
│ ├── Mistakes/ — errors to avoid YYYY-MM-DD_<slug>.md
│ ├── Context/ — project background YYYY-MM-DD_<slug>.md
│ ├── Sessions/ — one file per session YYYY-MM-DD_<session-id>.md
│ └── Index.md — chronological list, newest first
└── <another-project>/
└── ...
Where <project-name> is the last component of the current working directory (Path.cwd().name). Each project gets its own isolated subfolder.
After any manual edits in Obsidian:
agent-notes memory index
If you change a rule above (filename pattern, frontmatter field, time format), update both:
agent_notes/data/skills/obsidian-memory/SKILL.md)agent_notes/services/memory_backend.py — the CLI that implements themThe two must stay in sync. The skill is the canonical statement; the CLI is the enforcer.
The session note is updated on every meaningful state change. Specifically:
memory add ... session lead call that establishes the note.Skipping any of these makes the session note stale and breaks cross-session reconstruction. The Done Gate in global-claude.md enforces this for the lead; team agents follow the read-before-work side via the next section.
Whenever a session is active (Claude Code session ID detectable) and a non-session note is written, the CLI automatically appends a wikilink line to the session note's ## Linked notes section:
- [[<filename-stem>]] — <type> — <title>
This is done in the same agent-notes memory add call — no second call is needed. The operation is idempotent: if the link already exists, it is not duplicated.
The session note becomes the navigable hub for everything written during the session without any extra work from the agent.
Backend conditional: auto-linking only applies when the configured memory storage is Obsidian. On local storage, there are no wikilinks and this step is a no-op.
After every Claude Code ExitPlanMode invocation:
{{MEMORY_PATH}} at build time; if it resolves to a path under a vault, the storage is Obsidian.agent-notes memory add "<plan-title>" "<plan-body>" decision lead. The local plan file at ~/.claude/plans/<file>.md stays for harness compatibility.When mirrored, the new Decision participates in the Linking rule above — the active session note (if any) gets a wikilink to it.
Why mirror, not move: Claude Code's plan-mode requires the local file to exist (the harness reads it on resume and ExitPlanMode writes to it). The Decision note in Obsidian is the navigable canonical record; the local file is the harness's working copy.
| Choose Obsidian when... | Choose Wiki when... |
|---|---|
| You want category-based organization (Patterns, Decisions, Mistakes) | You want knowledge that compounds (sources → concepts → synthesis) |
| You need auto-linking between session notes and discoveries | You need the ingest → query → lint workflow |
| You browse notes visually in Obsidian with backlinks | You're building a team knowledge base |
| Your memory is about process (what worked, what failed) | Your memory is about domain knowledge (how things work) |
Process vs domain memory:
Obsidian storage focuses on process memory — tracking decisions, patterns, and mistakes across sessions. It answers "What did we learn?" and "Why did we choose this?"
Wiki storage focuses on domain memory — compiling source material into a structured, cross-referenced knowledge base that compounds over time. It answers "How does this work?" and "What are the facts?"
Any dispatched agent that needs context about the project's current state — recent decisions, patterns, mistakes, or what the lead has done so far — MUST read the Obsidian vault before starting work:
<vault>/Index.md to see what categories have content and which notes are most recent.<vault>/Sessions/<session-id>.md) for the latest progress.The vault path is substituted into agent prompts at build time as {{MEMORY_PATH}}. If {{MEMORY_PATH}} resolves to "disabled", skip this protocol — memory is not configured.
Agents do NOT need bash access for this; the vault is plain Markdown readable with the Read tool.
For ingesting external sources (URLs, files, folders), use the dedicated /ingest skill. It handles both wiki and obsidian backends automatically.
See: /ingest — adapts the Karpathy ingest → compile → cross-reference flow to whichever backend is active.