llm-wiki-brain

name	llm-wiki-brain
description	Filesystem-first LLM-Wiki + Obsidian agent memory system. Use when the agent needs to read, write, search, or maintain its persistent knowledge base. Provides brain search, lint, ingest-sessions, and build-index commands via the r0b0tlabbra1n package.
version	0.1.0
metadata	{"author":"@mr-r0b0t on X — r0b0tlab","hermes":{"tags":["memory","knowledge","obsidian","wiki","brain"],"category":"memory","min_hermes_version":"0.8.0"}}

LLM Wiki Brain — Hermes Agent Memory System

When to Use

Load this skill when the agent needs to:

• Search its persistent knowledge base for facts, procedures, or project context
• Record new durable information (decisions, failures, successes, project status)
• Check vault health (lint for broken links, stale items, secrets)
• Ingest Hermes sessions into the brain vault
• Navigate the vault structure

Requirements

• Python 3.11+
• r0b0tlabbra1n package installed (pip install r0b0tlabbra1n or pip install -e . from the repo)
• Brain vault initialized (brain init --vault PATH)
• Hermes state.db accessible for session ingest (read-only)

Quickstart

# Install
pip install r0b0tlabbra1n

# Initialize a brain
brain init --vault ~/my-brain

# Build the search index
brain build-index --vault ~/my-brain

# Search
brain search "how to deploy vLLM" --vault ~/my-brain

Core Workflows

1. Orient in a new session

When starting work, an agent should:

1. brain search "project status" --vault ~/my-brain — get current state
2. Read ~/my-brain/START_HERE.md and ~/my-brain/_agent/START_HERE.md
3. Read ~/my-brain/_agent/semantic/project-status.md for project context

2. Record a decision

# Create a decision page
python -c "
from r0b0tlabbra1n.vault.write_ops import safe_write_page
from r0b0tlabbra1n.models import PageFrontmatter, PageType, PageStatus, MemoryTier
from pathlib import Path

vault = Path.home() / 'my-brain'
page = vault / '_agent' / 'episodic' / 'decisions' / 'decision-slug.md'
fm = PageFrontmatter(
    title='Decision: Something Decided',
    type=PageType.DECISION,
    status=PageStatus.ACTIVE,
    tier=MemoryTier.L3_COLD,
    provenance='agent',
)
body = '''# Decision: Something Decided

**Context:** Why this decision was needed.

**Options considered:**
1. Option A
2. Option B

**Decision:** We chose X because...

**Impact:** What this affects going forward.
'''
safe_write_page(page, fm, body, provenance='agent')
print(f'Decision saved: {page}')
"

3. Search before acting

Before making changes, search for relevant past context:

brain search "project topic" --vault ~/my-brain --budget 2000

4. Lint the vault

Regularly check vault health:

brain lint --vault ~/my-brain

Expected output: zero ERRORs, zero WARNINGs. INFO (orphans) is fine.

5. Ingest new sessions

After significant work, ingest the session:

brain ingest-sessions --hermes-home ~/.hermes --vault ~/my-brain

Run brain build-index --vault ~/my-brain after ingestion.

Vault Structure

~/my-brain/
├── START_HERE.md          # Entry point
├── SCHEMA.md              # Structure and conventions
├── index.md               # Page index
├── log.md                 # Chronological activity log
├── _agent/                # Agent instructions and memories
│   ├── START_HERE.md
│   ├── operating-rules.md
│   ├── semantic/          # Durable facts
│   └── episodic/          # Session/event records
├── dashboards/            # Health and status dashboards
├── projects/              # Per-project pages
├── sessions/              # Session summaries
├── entities/              # People, orgs, tools
├── concepts/              # Technical concepts
└── _meta/                 # Indexes, manifests, config

Memory Tiers

• L1 HOT: Hermes built-in MEMORY.md (~2KB) — compact pointers only
• L2 WARM: Dashboards, project status, operating rules — auto-loaded
• L3 COLD: Session summaries, detailed pages — retrieved on demand
• L4 ARCHIVE: Raw captures, superseded pages — never auto-loaded

Promotion Rules

• Fact observed 3+ times → promote to L2 WARM
• Fact used 5+ retrievals → promote to L1 HOT as compact pointer
• Page stale > 30 days → demote to L3 COLD
• Page stale > 90 days → demote to L4 ARCHIVE

Secrets Policy

NEVER store secrets, API keys, tokens, or passwords in the vault. Store key names and config paths only. The secret scanner runs on every write.

Common Patterns

Find what you need:

brain search "keyword" --vault ~/my-brain

Check vault health:

brain lint --vault ~/my-brain

Update the index after changes:

brain build-index --vault ~/my-brain

Bring sessions into the brain:

brain ingest-sessions --vault ~/my-brain
brain build-index --vault ~/my-brain

Pitfalls

• Always run brain build-index after adding new pages or the search won't find them
• Session ingest is idempotent — safe to run repeatedly
• Use brain lint before committing changes to the vault
• The vault is filesystem-first: you can always read/write .md files directly
• Orphan pages are INFO-level only — not errors
• Secrets in the vault are hard-blocked by the secret scanner

Verification

After setup, verify:

brain init --vault /tmp/test-brain
brain lint --vault /tmp/test-brain
# Should show only INFO-level orphan messages, no ERRORs or WARNINGs