Obsidian Brain

A token-efficient session-memory layer for any project whose CLAUDE.md (preferred) or AGENTS.md references an Obsidian vault. Maintains a small index file (<slug>-brain.md) plus atomic notes for sessions, topics, and decisions, all colocated with the project's existing documentation in the vault.

Operates as a knowledge graph per project, complementing — never duplicating — the harness's built-in auto-memory (which captures user/feedback/project/reference snippets globally).

When to Use

• The current working directory has a CLAUDE.md or AGENTS.md that mentions an Obsidian vault, a vault path, or a folder named *Obsidian*.
• The user explicitly asks to load project memory, activate the brain, or runs /valarmindskills:obsidian-brain.
• The hook hooks/obsidian-brain/obsidian-brain-activate.js injected the activation system-reminder at session start.
• A relevant change was made (architecture, decision, refactor, gotcha) and you need to record it for the next session.

Do not use when

• No CLAUDE.md or AGENTS.md exists in the cwd or any ancestor up to the repo root.
• Neither file references an Obsidian vault — the regex obsidian|vault|Obsidian returns no match.
• The detected vault path resolves under a directory marked read-only by the project's CLAUDE.md (e.g., Notes/ in this repo). Refuse and report.
• The user asked for the built-in auto-memory types (user/feedback/project/reference) — that is a separate system handled by the harness.
• The task is a one-line edit, typo fix, or trivial dependency bump and the user has not asked for memory.

Prerequisites

If the CLI probe fails, the skill switches to fallback file-IO mode (Read / Write / Edit / Bash grep). All examples in this skill assume the CLI; substitutions are listed in references/READING_AND_SEARCHING.md and references/WRITING_RULES.md.

Probe once. Run obsidian --version 2>/dev/null exactly once per session, on the first turn the skill activates. Cache the result mentally as mode = cli | file and use that mode for the rest of the session. Do not re-probe per turn.

Phase 0 — Detection

Run, in order, until one resolves:

1. Read CLAUDE.md from the cwd. If absent, try AGENTS.md. If absent, walk up to a maximum of three ancestor directories.
2. Extract the first vault path from the file body. The path matches the regex below — relative (./, ../), absolute (/), or home-relative (~/), and must contain the literal substring Obsidian:

   (?:~\/|\.{1,2}\/|\/)[^"\s'`)]*Obsidian[^"\s'`)]*

The path terminates at whitespace, double-quote, single-quote, backtick, or close-paren. Expand ~ via the home directory if present. Resolve to an absolute path. If realpath fails or the path is read-only, abort with a one-line notice and exit.

1. From the resolved path, derive <project-slug> from the immediate child folder name (the leaf of the resolved path). Apply only lowercase + replace [^a-z0-9] with '-' then trim leading/trailing -. Do not apply camelCase splitting. Examples:

• Projetos/ValarMindSkills/ → slug valarmindskills → brain file valarmindskills-brain.md
• ~/MyVault/MyProject/ → slug myproject → brain file myproject-brain.md

This rule must match the hook's derivation in hooks/obsidian-brain/obsidian-brain-activate.js. If the hook output contradicts your computed path, trust the hook output.

1. Compute:

• vaultRoot = <resolved directory> (the project doc folder, not necessarily the entire vault — the brain lives next to the project's existing docs, not at the vault root)
• brainRoot = <vaultRoot>/brain/
• indexPath = <brainRoot>/<slug>-brain.md

If the resolved path looks like a vault root rather than a project folder (i.e., it contains a Projetos/ or similar subfolder), the user probably meant a child of it. Surface the ambiguity in one line and wait for confirmation before bootstrapping.

Record these signals in conversation context. Do not write to disk yet.

The full decision tree, including ambiguity resolution and conflicting CLAUDE.md/AGENTS.md, lives in references/SESSION_LIFECYCLE.md.

Phase 1 — Bootstrap (first run)

Trigger: <indexPath> does not exist.

Idempotency rule. If <indexPath> already exists on disk, skip Phase 1 entirely and go to Phase 2 — never re-bootstrap, never overwrite the index. Re-bootstrap is only allowed if the user explicitly says so (e.g., "wipe the brain and start over"), and even then only after a one-question confirmation.

[!important] When to ask the bootstrap question Phase 1 is eager-ask, lazy-write — the question fires as the agent's first user-facing sentence; the actual mkdir + seed only happens after the user says yes. The timing rule:

>

- Default (every session, including auto mode) — ask the question immediately after Phase 0 succeeds, as the first user-facing sentence of the next response. Auto mode does not suppress the question: a single y/n is one keystroke, not a real interruption, and Phase 1 has no other reliable trigger that does not depend on the brain already existing (Phase 3 needs the index to know what to write; Phase 4 needs the index to sync). - Mid-task exception (rare) — if another skill is already mid-question (an actual interactive prompt is on screen waiting for input), defer only until that prompt is answered, then ask. Do not defer to "first natural pause", "end-of-task report", or any other vague handoff — those are the failure modes that left brains uncreated. Record the deferral as a single one-shot intent so it survives /compact. - Phase 3 race guard — if any Phase 3 write trigger (new session note, topic, decision) fires before Phase 1 has been answered, stop and ask now. Never write to brain/ without an existing index. - Phase 4 entry guard — if any end-of-session signal fires before Phase 1 was answered, ask now before running any Phase 4 action. Phase 4 is gated on Phase 1. - Re-injection idempotency — the SessionStart hook re-injects the activation digest after /compact and across new sessions. If the digest fires while Phase 1 is still pending, resume the same intent (do not re-defer indefinitely, do not ask twice). Treat the deferral state as a single one-shot until answered or the user explicitly refuses. - Refusal is sticky for the session — once the user says no, do not re-prompt; mark the brain as opted-out for the rest of the session and respect that across re-injections.

Steps:

1. Confirm with the user before any write: "Bootstrap obsidian-brain at <indexPath>? This creates brain/, three subdirectories, and seeds the index. Proceed?" Wait for explicit yes. Choose the moment per the timing rule above.
2. Create the directory tree. Phase 1 prefers the shell-IO path because directory creation and bulk file seeding are more reliably done with mkdir/Write than via the CLI's create subcommand (which the local @obsidian-cli skill exemplifies only with name=, not path=). Use:

   mkdir -p "<brainRoot>/sessions" "<brainRoot>/topics" "<brainRoot>/decisions"
   touch "<brainRoot>/sessions/.gitkeep" "<brainRoot>/topics/.gitkeep" "<brainRoot>/decisions/.gitkeep"
   # then use the Write tool to seed <indexPath> from the template in references/STRUCTURE.md.

If you have a strong reason to use the CLI for the index seed instead (e.g., to integrate with vault watchers), confirm syntax first: obsidian help create then attempt obsidian create path="brain/<slug>-brain.md" content="<seeded index>" silent. If that fails, fall back to Write immediately — never guess flags.

1. Seed <indexPath> from the index template in references/STRUCTURE.md. The seed must satisfy the token economy budget: ≤500 tokens total, ≤150 tokens in the > [!summary] "Critical facts" callout.

1. Do not modify any file outside `brain/`. Brain is an isolated graph — the project's main MOC must not be edited to point at the brain, and the brain must not contain wikilinks ([[...]]) or embeds (![[...]]) to any file outside brain/ (project docs, Notes/, Daily Notes, attachments, or any other vault folder). Phase 1 touches only the brain/ subtree and its index. External entities are referenced by plain text or backticks only.

1. Output the bootstrap report (see Output format).

If the user refuses bootstrap, exit silently — do not retry within the session.

Phase 2 — Load (read strategy)

Trigger: <indexPath> exists.

[!info] Route the prompt before reading Before loading anything, classify the user's question into one of three buckets — the answer determines which subtree of the vault to consult:

>

| Bucket | Examples | Where the answer lives | Phase to follow | | :--- | :--- | :--- | :--- | | Brain-history | "what was decided about X", "when did we discuss Y", "open todos", "recent activity" | brain/sessions, brain/topics, brain/decisions | Phase 2 (this section) | | General-doc | "how does X work", "what does skill Y do", "where is the architecture for Z", "show me the MOC" | Project docs (<vaultRoot>/<MOC>.md, Skills/, Arquitetura.md, Manual de Uso/, Technical Design/) | Phase 2.5 | | Mixed | "explain auth and what we decided about it" | Both — but lead with the general doc for substance, then surface brain context | Phase 2.5 then Phase 2 |

>

Default to Phase 2.5 if the question is about how the system works. Default to Phase 2 only when the question is about what happened in past sessions. When unclear, run Phase 2.5 first; the brain is cheap to consult after.

1. Read only the index at session start (≤500 tokens). Do not preemptively load sessions, topics, or decisions.
2. From the index, identify the wikilinks relevant to the current user prompt — match by topic name, tag, or recent-session date.
3. Lazy-load only matched targets:

• For a topic: obsidian read file="<topic>" (CLI, exemplified) or Read on the absolute path (fallback).
• For a tag query: obsidian search tag="<tag>" (CLI, verify with `obsidian help search` first) or grep -lrn 'tags:.*<tag>' <brainRoot> (fallback).
• For a property query: obsidian search property="type=brain-decision" (CLI, verify) or grep -lrn '^type: brain-decision' <brainRoot> (fallback).

1. Stop reading the moment the relevant context is in hand. Do not "read everything just in case."
2. Critical-facts callout in the index is the only block you may quote verbatim into your reasoning. Everything else is a wikilink to follow on demand.

Detailed strategy and the canonical CLI command catalog: references/READING_AND_SEARCHING.md.

Phase 2.5 — Searching general project docs

Trigger: the user's question is about how the system works (general-doc bucket above) or the brain index has no matching wikilink.

The brain captures what happened; the project's main vault docs capture how it works. When the answer lives in the main docs, you must search them with the CLI before reaching for Read/grep.

1. Use the cached `mode` from Phase 2. If mode = cli, run the CLI recipes below. If mode = file, fall back to grep/find/Read.
2. Scope the search to the project doc folder, not the entire vault:

   obsidian search query="<term>" path="<vaultRoot>"     # verify with `obsidian help search` first

<vaultRoot> here is the same path computed in Phase 0 (e.g., Projetos/ValarMindSkills/).

1. Filter by tag or property when the user named a category — tag="skill", tag="lang/go", property="type=skill". Combine with total to count first:

   obsidian search tag="skill/segurança-api" total

1. Read the top match with obsidian read file="<note-name>" silent. Aliases resolve automatically — prefer file= over path= unless there is ambiguity.
2. Optionally check backlinks to surface related sessions in the brain:

   obsidian backlinks file="<note-name>"

1. Quote sparingly. Pull only the one or two passages that answer the question. Link to the doc with [[<note>]] when writing back to the brain — do not duplicate the doc into a topic note.

Full intent→command recipe table, fallback chain, and Notes/-read-only reminder live in references/READING_AND_SEARCHING.md § Searching general project docs. Canonical CLI catalog (used by every other agent too) is in @obsidian-cli.

Always probe @obsidian-cli first; only fall back to Read/grep when the cached mode = file. Aliases, wikilinks, and the vault index are invisible to plain grep.

Phase 3 — Operate (write strategy)

Write to the brain only when at least one of these triggers fires:

Always:

• Dedupe first — obsidian search query="<key term>" or obsidian tags list before creating a new topic/decision. If a near-duplicate exists, append/refine it instead of creating a new note.
• Bullets > prose — every entry begins with a one-sentence summary line, then bullets, then wikilinks. No paragraphs.
• Internal wikilinks ≥ 2 per non-trivial entry — the entry must connect into the brain graph via wikilinks to other notes under brain/ only. Wikilinks ([[...]]) and embeds (![[...]]) across the brain/ boundary — to project docs, Notes/, Daily Notes, attachments, or any other vault file — are forbidden. Reference external entities by plain text or backticks (` Arquitetura.md , prompt-engineering `). See Constraints.
• Update properties surgically — obsidian property:set name="updated" value="YYYY-MM-DD" file="<note>" instead of rewriting the frontmatter.
• Refuse `Notes/` — if any target path falls under Notes/ (or any directory the project CLAUDE.md marks read-only), abort the write and notify the user.

Templates per note type, dedupe heuristics, ADR triggers, and atomic-note rules: references/WRITING_RULES.md.

Phase 4 — End-of-session sync

When the user signals end-of-session ("/clear", "obrigado", "encerrar", actual session close, or after a substantial finished task), perform this sequence at most once per session:

1. Phase 1 entry guard. If <indexPath> is missing AND the session had ≥1 substantive turn (anything beyond "hi"/typo fix), surface the Phase 1 bootstrap question now, before any other Phase 4 action. Block the rest of Phase 4 until the user answers. If the user refuses, exit Phase 4 silently — there is nothing to sync. This guard catches the case where the SessionStart digest was emitted but the agent failed to ask the question earlier in the session.
2. Append session summary as a > [!summary] callout to the active session note in brain/sessions/. ≤30 lines, ≤6 bullets.
3. Update properties of touched topics and the index (updated: YYYY-MM-DD) via surgical CLI calls.
4. Refresh "Recent sessions" in the index — wikilink to today's session note at the top, drop entries older than 10. If buildup exceeds 10, suggest the dedicated synthesize command.
5. Suggest synthesis when 3+ recent sessions touched the same topic without a corresponding topics/<topic>.md note. One line: "Topic X appeared in 3 recent sessions; want me to synthesize a topic note?"

Do not write a session note if the session produced nothing worth remembering (informational chat, trivial fix, no architectural insight). Bias toward fewer notes.

Phase 5 — Suggest updating main docs

The brain captures what happened; the project's main vault docs (Projetos/<project>/*.md MOCs, Architecture notes, Skill notes) capture how the system works. The two are complementary, not redundant.

After any change that alters architecture, conventions, public interfaces, or skill behavior:

1. Identify the canonical doc affected (typically Arquitetura.md, Manual de Uso.md, Skills/<slug>/<slug>.md, or the project MOC).
2. Surface a one-line suggestion: "Consider updating <doc> to reflect <change>. Want me to draft the diff?"
3. If the user agrees, hand off to @obsidian-markdown for syntax and to @obsidian-cli for the actual write.

Do not update main docs unilaterally. The user reviews the diff first. The Phase 5 suggestion is prose only — never create a wikilink ([[...]]) or embed (![[...]]) from any brain note to a file outside brain/, and never edit the project MOC (or any other non-brain file) to wikilink/embed to the brain. Brain stays out of the project documentation graph in both directions; references across the boundary use plain text or backticks only.

Skills it delegates to

This skill orchestrates the strategy (what to write where, when to read what); the delegated skills handle the mechanism.

Persistence

Default mode: on. The SessionStart hook auto-activates per session whenever a vault is detected via CLAUDE.md or AGENTS.md. With no vault detected, the hook silently clears the active flag and the statusline badge dims to cinza — the skill produces no further effect that session.

The skill does not inject reinforcement into every UserPromptSubmit — the SessionStart digest is sufficient. Adding per-turn context would defeat the token-economy goal.

Constraints

• Never write under any directory the project's CLAUDE.md marks read-only (Notes/ in this repo).
• Never duplicate auto-memory built-in content (user preferences, global feedback, profile facts).
• Never load more than the index at session start. Lazy-load only.
• Never exceed 500 tokens in the index file. Critical-facts callout ≤150 tokens.
• Never write prose paragraphs in brain files — bullets, callouts, wikilinks only.
• Never rewrite a note when only a property changed — use surgical property update.
• Never edit an ADR after it has been written — supersede via a new file with status: superseded.
• Never include wikilinks ([[...]]) or embeds (![[...]]) from any brain note (sessions, topics, decisions, or the index) to any file outside `brain/`. This covers project docs, MOCs, Arquitetura.md, Manual de Uso.md, Skills/<slug>/*.md, Notes/, Daily Notes, attachments, .base files, and every other vault folder. The brain/ subtree is the only valid wikilink target inside brain notes — files generated by this skill must reference external entities exclusively by plain text or backticks (` prompt-engineering , Arquitetura.md `). If a brain note needs to point at a non-brain file, write the file name in backticks, not as a wikilink.
• Never modify the project MOC, Arquitetura.md, or any other project doc to add a wikilink ([[...]]) or embed (![[...]]) to the brain. Brain is a closed graph; isolation is bidirectional — nothing inside brain/ links out, nothing outside brain/ links in.
• Must refuse the write and surface the violation if any draft session/topic/decision note contains a wikilink or embed whose target resolves outside brain/. Strip the offending link (replace with backticks) before committing, or abort the write and ask the user how to phrase the reference.
• Must prefer @obsidian-cli for every I/O operation for commands marked exemplified in references/READING_AND_SEARCHING.md. For commands marked verify, run obsidian help <subcommand> once before first use; on failure, drop to file-IO without retrying with guessed flags.
• Must dedupe before creating any topic or decision note.
• Must include a > [!summary] callout as the first content block of every session note.
• Must ask the user before bootstrapping the brain (one explicit yes/no, once per session).
• Must suggest updating the main docs after relevant changes — never update them unilaterally.
• Must treat the slug derivation rule (lowercase + [^a-z0-9]→'-', no camelCase) as authoritative; if it produces a name that conflicts with an existing file, surface the conflict instead of inventing an alternative.

Output format

After any operation, report exactly:

obsidian-brain: <action>
  index: <indexPath> (mode: cli|file)
  wrote: <files written, paths relative to brain/>
  read: <files loaded, count>

Suggested next: <one line — "synthesize topic X" / "update Arquitetura.md" / nothing>

Omit the report when the operation was a silent skip (no detection match, user refused bootstrap, nothing worth writing).

Example invocations

• (automatic) Hook injects OBSIDIAN-BRAIN ACTIVE at session start when CLAUDE.md / AGENTS.md mentions a vault.
• "ative o brain do projeto"
• "carregue a memória do projeto antes de responder"
• "load project memory"
• "/valarmindskills:obsidian-brain"
• "/valarmindskills:obsidian-brain on" — re-enable after a previous off in the same session.
• "/valarmindskills:obsidian-brain off" — disable for this session.
• "stop obsidian-brain", "desativar obsidian-brain" — natural-language disable.
• "/valarmindskills:obsidian-brain synthesize" — consolidates recent sessions into topic notes.

References

• SESSION_LIFECYCLE — detection decision tree, first-run vs subsequent runs, end-of-session triggers.
• STRUCTURE — exact brain/ layout, frontmatter for every note type, naming conventions, callouts, templates.
• READING_AND_SEARCHING — lazy-load discipline, canonical CLI commands, fallback chain, brain.base on-demand.
• WRITING_RULES — atomic notes, dedupe via CLI, ADR triggers, append-only sessions, atomic-rewrite topics, fallback substitutions.