brain

name	brain
description	Persistent cross-session memory in a configured Obsidian vault owned by the agent. Load on turn 1 of EVERY session before responding. Also load when durable facts, preferences, links, screenshots, project references, gotchas, or improvement items surface.

Brain — Cross-Session Memory via Obsidian

Your long-term memory lives in the configured Obsidian vault. The vault is the handoff from prior-you to future-you.

Configuration

The vault name and path are deployment-specific. Read them from references/brain-config.md at the start of each session and substitute them wherever this skill or its guides reference <brain-vault-name> or <brain-vault-path>.

If references/brain-config.md is missing, stop before doing any vault work and ask the user to run:

npx tsx scripts/configure-brain.ts --vault-name "vault-name" --vault-path "/absolute/path/to/vault"

After substitution, the rest of this skill and its */GUIDE.md files apply unchanged.

The primary purpose is project understanding: mental models, decisions, gotchas, conventions, links, screenshots, diagrams, external references, and enough context that a fresh agent can resume without re-discovery. Preferences, tool notes, and backlog items support that purpose.

This brain is yours, not the user's. The user should not need to manage it. Write proactively when future-you will benefit.

First Rules

• Turn 1 always loads startup. Read startup/GUIDE.md and index.md before substantive work. Do not ask permission.
• Meta turns are allowed to be explicit. If the user asks about the brain/vault/memory system, still orient first, but discussing the orientation and process openly is fine.
• Project slugs are lowercase-kebab-case. Normalize repo/folder basename: lowercase, replace non-alphanumeric runs with -, strip edges. Example: my.project.core -> my-project-core.
• Path-qualify ambiguous wikilinks. If a basename exists in multiple places (improvements.md, gotchas.md, decisions.md, architecture.md), link with a path and alias: [[projects/<slug>/improvements|improvements]].
• Read before writing. Search existing notes first; update in place instead of creating near-duplicates.
• Verified facts only. Never store guesses as facts. If uncertain, mark uncertainty or ask.
• No secrets. Do not store credentials, API keys, tokens, private env values, or sensitive customer data.

Vault Layout

<brain-vault-name>/
├── index.md                    # Lightweight TOC + project status
├── improvements.md             # Cross-cutting backlog only
├── person-*.md                 # People and preferences
├── pref-*.md                   # Global preferences not tied to a person
├── tool-*.md                   # Tool usage patterns
├── codestyle-*.md              # Language idioms
├── decision-*.md               # Cross-project decisions
├── gotcha-*.md                 # Reusable traps
└── projects/<slug>/
    ├── _<slug>.md              # Required project entry point and living mental model
    ├── conventions.md          # Optional; create when useful
    ├── decisions.md            # Optional; create when useful
    ├── gotchas.md              # Optional; create when useful
    ├── improvements.md         # Optional; create when there are project backlog items
    ├── <topic>.md              # Optional deep topic notes
    └── <attachments>           # Screenshots, diagrams, PDFs, exports, reference files

Only projects/<slug>/_<slug>.md is mandatory for a project. Other companion files are as-needed; empty stubs are noise.

index.md should stay lightweight. It links to projects and important root notes; do not duplicate detailed project summaries that belong in _<slug>.md.

The updated: frontmatter field is useful but advisory. If stale, trust filesystem mtime and note content over the date.

Project Entry Point

Use projects/<slug>/_<slug>.md as both TOC and mental model. Keep it current when understanding changes.

---
created: YYYY-MM-DD
updated: YYYY-MM-DD
kind: service|library|app|tool|monorepo|other
stack: [language, runtime, key-frameworks]
status: active|dormant|archived
depends_on: []
used_by: []
---

# <project-slug>

## Purpose
## Architecture
## Tech stack
## Layout
## Relationships
## Target state
## Open questions
## Notes for future-you
## Companion notes and references

#project/<slug>

Every project-scoped markdown note must end with #project/<slug>.

What To Remember

Capture normal language, not just explicit commands. The user should not need magic phrases.

• Person/preferences: The user says they like/dislike a behavior, corrects interaction style, repeats a workflow habit, or expresses frustration. Update person-user.md or pref-*.
• Project facts: purpose, architecture, deploy, data flow, ownership, target state, current constraints. Update _<slug>.md or a topic note.
• Decisions: dated choice plus reasoning. Append to decisions.md or root decision-*.
• Gotchas: traps, failed approaches, doc/code drift, non-obvious workarounds. Write gotchas.md or root gotcha-*.
• Tool patterns: reusable CLI/API/GraphQL/auth commands. Write tool-*.
• External references: URLs, Slack/Teams threads, Figma links, docs, screenshots, PDFs, diagrams, vendor references, issue/PR links. Store or link them where future-you will look.
• Improvements: fixable friction that is not the current task. Append one line to root or project improvements.md.

Counterexample: "I'm going to run tests now" is not durable. "I prefer just test as the canonical test command" is durable.

External References And Attachments

Project references are first-class memory. When a link, screenshot, document, or external artifact explains a project, preserve it in the project folder.

• Links: use normal markdown links with context: [dashboard runbook](https://...) — source of truth for prod restart steps as of YYYY-MM-DD.
• Screenshots/images: if the file is available locally, copy or move it into projects/<slug>/ when safe, then embed it from the note with ![[filename.png]]. Use descriptive lowercase-kebab filenames like login-error-2026-04-29.png.
• PDFs/docs/diagrams: keep small, load-bearing files next to the note that uses them and link/embed via Obsidian syntax.
• External-only assets: do not mirror large or sensitive assets. Link them with a short explanation, owner/source, and why future-you cares.
• Reference index: if a project accumulates several external artifacts, create projects/<slug>/references.md and link it from _<slug>.md.
• Safety: never save secrets, private credentials, sensitive customer data, or giant binary dumps. If unsure, ask the user.

Capture Workflow

1. Classify scope: root or project.
2. Search existing notes with Grep / Glob before writing.
3. Edit the existing note when possible; create a new note only when the topic earns one.
4. Link new project notes from _<slug>.md; link important root notes from index.md.
5. If the user explicitly asked to remember something, confirm in one line where it was saved. Otherwise silent capture is fine.

Common target mapping:

• User preference -> person-user.md
• Global preference -> pref-<slug>.md
• Tool pattern -> tool-<slug>.md
• Reusable trap -> gotcha-<slug>.md
• Project mental model -> projects/<slug>/_<slug>.md
• Project references -> projects/<slug>/references.md or the relevant topic note
• Project decision -> projects/<slug>/decisions.md
• Project gotcha -> projects/<slug>/gotchas.md
• Project improvement -> projects/<slug>/improvements.md
• Cross-cutting improvement -> root improvements.md

Extract Workflow

Run at wrap-up, after milestones, before compaction, and after long multi-step stretches.

Checkpoint questions:

• Did the user express a preference, correction, communication style, workflow habit, or recurring frustration? Update person-user.md.
• Did the project mental model change? Rewrite _<slug>.md in place; do not append stale caveats forever.
• Were links, screenshots, docs, diagrams, issues, PRs, or external references used as evidence? Store/link them under the project.
• Did we learn a reusable tool pattern? Update/create tool-*.
• Did new fixable friction appear? Log it in the appropriate improvements.md.
• Is the total open backlog roughly over 10 items? Tell the user the backlog is growing and ask whether to run an improvement pass.

Report briefly only when useful: what was extracted and where.

Improvement Workflow

Capture improvements opportunistically, but do not fix backlog items unless the user asks for a pass.

Format:

- [ ] YYYY-MM-DD (low|med|high) — [[target]] — description. (proposed fix, if obvious)

During an explicit pass, surface grouped open items and ask what to work through. Fix one item at a time. When fixed, move it to ## Fixed (recent) in the same file.

Consolidation Workflow

Run when the user asks for cleanup or obvious duplication/staleness appears.

• Merge duplicates into the canonical note.
• Delete only with user confirmation.
• Rewrite stale project notes in place when architecture materially changed.
• Backfill links and missing #project/<slug> tags.
• Do not create empty project companion stubs.

Tools

Prefer native file tools for CRUD: Read, Grep, Glob, and the available file-editing primitive (apply_patch, Edit, or Write, depending on the host toolset).

Use the configured Obsidian CLI with vault=<brain-vault-name> for indexed search, backlinks, unresolved links, tag analytics, graph-aware moves, daily notes, and plugin/app operations. Obsidian must be running for CLI commands.

Useful CLI forms:

• obsidian vault=<brain-vault-name> search query="terms"
• obsidian vault=<brain-vault-name> backlinks path=projects/<slug>/_<slug>.md
• obsidian vault=<brain-vault-name> unresolved
• obsidian vault=<brain-vault-name> tags counts sort=count
• obsidian vault=<brain-vault-name> move path=<old> to=<new>

Argument names are path= for file operations and query= for searches; do not use file= for project notes.

When Not To Write

• Ephemeral session state.
• Repo-owned documentation that should live in the repo.
• Unverified assumptions.
• Secrets or sensitive data.
• Large binary archives or screenshots with sensitive contents.