Skip to content

Knowledge Vault

The knowledge vault is an Obsidian-based persistent memory layer at .ai-brain/. It stores goals, decisions, reports, and reference knowledge that survives across conversations. Agents read from and write to the vault through the brain skill, and semantic search is provided by the obsidian-rag MCP server.

Vault Structure

.ai-brain/
├── 00-inbox/          # Landing zone — daily briefings and new notes
├── 01-goals/          # Quarterly objectives with key results
├── 02-projects/       # Project monitoring configs (repos, briefing checks)
├── 03-knowledge/      # Durable reference material (OSDU docs, architecture)
├── 04-reports/        # Generated reports (QA, dependencies, incidents)
├── attachments/       # Images, diagrams, embedded files
└── templates/         # Note templates (read-only)
Folder Contents Lifecycle
00-inbox Daily briefing notes, unprocessed items Created daily, promoted or archived
01-goals Quarterly OKRs, key results tracking Updated weekly/monthly
02-projects Repo lists, monitoring configs, briefing rules Updated when project scope changes
03-knowledge OSDU platform docs, architecture decisions, standards Long-lived, updated when sources change
04-reports QA reports, dependency scans, incident RCAs Created per event, reference only

Note Types and Templates

Templates live in .ai-brain/templates/ and define the structure for each note type:

Template Used For Created By
daily-note.md Morning briefing output Wingman (briefing skill)
decision-record.md Architecture Decision Records Any agent (brain skill)
decision.md Lightweight decisions (non-ADR) Any agent (brain skill)
qa-report.md Test reliability analysis OSDU agent
dependency-report.md Dependency scan results OSDU agent
incident-report.md Outage/issue RCA Any agent
goal.md Quarterly objective + key results Wingman
architecture-note.md System design documentation Any agent (learn skill)
report-note.md General structured report Any agent
meeting-note.md Meeting minutes and action items Wingman

Write Discipline

Agents follow strict rules about what goes into the vault:

Write To Vault Keep In Chat
Decisions with long-term impact Ephemeral debugging output
Reusable knowledge (distilled from sources) Raw API responses
Reports that will be referenced later One-time analysis
Goals and key results Conversation-specific context
Daily briefing summaries Intermediate work products

The brain skill enforces these boundaries. Agents never write directly to vault files — they always go through the brain skill protocol, which validates content and applies the correct template.

MCP Integration

The obsidian-rag MCP server provides semantic search over the vault. Configured in .copilot/mcp-config.json:

{
  "mcpServers": {
    "obsidian-rag": {
      "command": "bash",
      "args": ["-c", "source .ai-wingman/.env && obsidian-rag serve --vault .ai-brain"],
      "env": {}
    }
  }
}

Available Tools

MCP Tool Purpose
search_vault Semantic search across all vault notes
get_note Retrieve a specific note by path
search_by_tag Find notes with specific Obsidian tags
get_related Find notes related to a given note
list_recent List recently modified notes
search_with_reasoning Semantic search with explanation of relevance
explore_connected_conclusions Traverse linked conclusions across notes
get_conclusion_trace Trace the reasoning chain for a conclusion
index_status Check vault indexing status

How Agents Use MCP

Agents query the vault before starting work to gather relevant context:

  1. Wingman searches for goals and prior decisions before making recommendations
  2. OSDU agent searches for past QA reports before analyzing test results
  3. Base copilot searches for architecture decisions before suggesting changes

Knowledge Flow

Vault Data Flow

Knowledge moves through the system in a defined pipeline:

Stage Action Tool
Discover Identify external sources (GitLab wikis, docs, URLs) learn skill
Fetch Retrieve raw content learn skill + web tools
Distill Extract key insights, remove noise learn skill
Write Create vault note with proper template brain skill
Index MCP server re-indexes vault content obsidian-rag (automatic)
Retrieve Agents search vault for relevant context MCP tools

Example: Learning from OSDU Documentation

User: "Learn about the OSDU Partition service from the wiki"

1. learn skill → fetches GitLab wiki pages for os-partition
2. learn skill → distills: API endpoints, data model, dependencies
3. brain skill → writes to 03-knowledge/osdu-partition.md
4. obsidian-rag → re-indexes, note available for future searches
5. Next conversation → agent finds partition context automatically

Vault Maintenance

The vault is designed to be low-maintenance:

  • Scaffolding is version-controlled — folder structure and templates live in the repo
  • Content is gitignored — actual notes are local to each developer's machine
  • Reindexing is automated — MCP server watches for file changes
  • Stale notes decay naturally — daily notes accumulate but don't need cleanup
  • Knowledge notes are updated — the learn skill updates existing notes when sources change

See also