Skip to main content
Rowboat automatically builds an Obsidian-style knowledge graph from your emails, meetings, and voice memos. This living knowledge base tracks people, organizations, projects, and topics you interact with.

​
How It Works

The knowledge graph system continuously monitors your data sources and extracts meaningful entities to create interconnected notes.

​
Processing Flow

1

Sync runs

Data sources (Gmail, Fireflies, Granola) fetch new or updated content and save as markdown files in ~/.rowboat/
2

Change detection

The system uses a hybrid mtime + hash approach to identify which files need processing:
  • Quick check: Compare file modification time (mtime)
  • Verification: If mtime changed, compute content hash to confirm actual content changes
3

Batch processing

New or changed files are processed in batches of 10 by the note_creation agent
4

Entity extraction

The agent extracts entities and creates/updates notes in ~/.rowboat/knowledge/
5

State tracking

Progress is saved incrementally after each successful batch to ~/.rowboat/knowledge_graph_state.json

​
Knowledge Structure

Rowboat organizes your knowledge into four primary entity types:

People

Contacts you interact with via email or meetingsFields tracked:
  • Name and email
  • Organization and role
  • Aliases for name variations
  • Communication history

Organizations

Companies and institutions mentioned in your communicationsFields tracked:
  • Organization name
  • Domain (e.g., example.com)
  • Aliases (e.g., “ACME” vs “ACME Corp”)
  • Related people and projects

Projects

Initiatives, deals, or ongoing work mentioned across conversationsFields tracked:
  • Project name
  • Status (active, completed, etc.)
  • Aliases
  • Related people and organizations

Topics

Concepts and subject matter that appear in your workFields tracked:
  • Topic name
  • Keywords for matching
  • Aliases
  • Related entities

​
Note Creation from Sources

​
From Emails

When processing email threads, Rowboat:
  1. Converts the thread to markdown format
  2. Extracts sender information, subject, and body content
  3. Identifies people, organizations, and topics mentioned
  4. Creates or updates corresponding notes with context
Email processing behavior is controlled by strictness levels - see the section below for details.

​
From Meetings

Meeting transcripts (from Fireflies or Granola) are processed similarly:
  1. Transcript is converted to structured markdown
  2. Participants are extracted and linked to People notes
  3. Topics discussed are identified and tracked
  4. Action items and decisions are preserved
Meetings always create notes at all strictness levels, as they represent significant interactions.

​
From Voice Memos

Voice memos recorded in Rowboat:
  1. Are transcribed and saved to ~/.rowboat/knowledge/Voice Memos/
  2. Undergo entity extraction like other sources
  3. Can reference existing people, projects, or topics
  4. Are organized by date for easy retrieval

​
Strictness Levels

Rowboat supports three strictness levels that control how aggressively notes are created from emails. This prevents your knowledge base from getting overwhelmed with noise.
Philosophy: “Meetings create notes. Emails enrich them.”
  • Emails never create new notes (only meetings do)
  • Emails can only update existing notes for people you’ve already met
  • Exception: Warm intros from known contacts can create notes
  • Best for: Users who get lots of emails and want minimal noise
What gets filtered:
  • Mass newsletters ✓
  • Automated/system emails ✓
  • Consumer services (Amazon, Netflix) ✓
  • Generic cold sales ✓
  • Recruiters ✓
  • Support reps ✓
  • Personalized business emails ✓
What creates notes:
  • Meetings ✓
  • Warm intros from existing contacts ✓

​
Auto-Configuration

On first run, Rowboat analyzes your emails and recommends a strictness level:
  • More than 100 human senders → Recommends High (avoid overload)
  • 50-100 senders → Recommends Medium (balanced)
  • More than 50% consumer services → Recommends Medium (filter noise)
  • Less than 30 senders → Recommends Low (comprehensive capture is manageable)
Strictness is configured in ~/.rowboat/config/note_creation.json and can be changed at any time.

​
Knowledge Index

Rowboat maintains a fast in-memory index of all your notes to help the AI quickly resolve entities without searching. 
{
  "people": [
    {
      "file": "People/John Doe.md",
      "name": "John Doe",
      "email": "[email protected]",
      "organization": "Acme Corp",
      "role": "Engineering Manager",
      "aliases": ["John", "JD"]
    }
  ],
  "organizations": [
    {
      "file": "Organizations/Acme Corp.md",
      "name": "Acme Corp",
      "domain": "acme.com",
      "aliases": ["ACME", "Acme"]
    }
  ],
  "projects": [...],
  "topics": [...],
  "buildTime": "2026-02-28T10:30:00.000Z"
}
The index is rebuilt before each processing batch to include notes from previous batches, ensuring entities are properly merged.

​
State Management

The knowledge graph maintains state in ~/.rowboat/knowledge_graph_state.json: 
{
  "processedFiles": {
    "/path/to/file.md": {
      "mtime": "2026-01-07T10:30:00.000Z",
      "hash": "a3f5e9d2c8b1...",
      "lastProcessed": "2026-01-07T10:35:00.000Z"
    }
  },
  "lastBuildTime": "2026-01-07T10:35:00.000Z"
}
This ensures:
  • Only new or changed files are processed
  • Progress is preserved even if processing is interrupted
  • You can force reprocessing by deleting the state file

​
Source Code Reference

Key implementation files:
  • apps/x/packages/core/src/knowledge/build_graph.ts - Main orchestrator
  • apps/x/packages/core/src/knowledge/graph_state.ts - State management
  • apps/x/packages/core/src/knowledge/knowledge_index.ts - Index builder
  • apps/x/packages/core/src/knowledge/sync_gmail.ts - Gmail integration
  • apps/x/packages/core/src/knowledge/sync_fireflies.ts - Meeting transcripts

Build docs developers (and LLMs) love

Get started for freeTalk to us