Open Brain:
Email Capture

Nate's guide cracked the daily capture habit. You type a thought into Slack, it gets embedded and classified in five seconds, and any AI you use can search it by meaning. That's genuinely transformative — but it captures what you decide to capture.

Your email is different. It's thinking you already did. Every email you write is a decision, a position, a relationship update, or a problem you solved. You wrote it, you sent it, and it immediately became invisible to your AI tools.

When we looked at 30 days of sent mail — project updates, client advice, personal advocacy, community work — it was 153 distinct thoughts that an AI couldn't access until now. Long emails explaining complex situations. Narrative summaries of ongoing projects. Strategic advice to colleagues. All of it, in your voice, about the things that actually matter to you.

How We Built It

This pipeline was built in a single session with Dr. Brian — an AI agent running in Cursor. We're sharing the story because the hard parts weren't where we expected them.

What we thought would be hard: Gmail OAuth. Getting permission to read someone's email requires Google Cloud Console setup, consent screens, scopes, token refresh. It's genuinely involved.

What was actually hard:

Credential Tracker

You'll generate one new set of credentials. Add these to your tracker alongside the ones from Nate's guide:

GMAIL (Step 2)
  Google Cloud Project:   ____________
  OAuth Client ID:        ____________
  OAuth Client Secret:    ____________
  credentials.json saved: yes / no

1. Get the code

   Clone the repo or pull the latest if you already have it.

   git clone https://github.com/MonkeyRun-com/monkeyrun-open-brain.git
   cd monkeyrun-open-brain
   # or if you already have it:
   git pull

2. Create Google Cloud OAuth credentials

   Go to console.cloud.google.com → New Project → "Open Brain". Enable the Gmail API. Configure the OAuth consent screen (External, add your Gmail as a test user, add the gmail.readonly scope). Create an OAuth Client ID (Desktop app type). Download the JSON and save it as scripts/credentials.json.

3. Run the database migration

   In the Supabase dashboard → SQL Editor, paste and run:

   -- Add chunking support
   ALTER TABLE thoughts
     ADD COLUMN IF NOT EXISTS parent_id uuid REFERENCES thoughts(id) ON DELETE CASCADE,
     ADD COLUMN IF NOT EXISTS chunk_index integer;
   
   CREATE INDEX IF NOT EXISTS thoughts_parent_id ON thoughts (parent_id)
     WHERE parent_id IS NOT NULL;
   
   -- RPC to bypass PostgREST schema cache
   CREATE OR REPLACE FUNCTION insert_thought(
     p_content text, p_embedding vector(1536),
     p_metadata jsonb, p_parent_id uuid DEFAULT NULL,
     p_chunk_index integer DEFAULT NULL
   ) RETURNS uuid LANGUAGE plpgsql AS $$
   DECLARE new_id uuid;
   BEGIN
     INSERT INTO thoughts (content, embedding, metadata, parent_id, chunk_index)
     VALUES (p_content, p_embedding, p_metadata, p_parent_id, p_chunk_index)
     RETURNING id INTO new_id;
     RETURN new_id;
   END; $$;

4. Deploy the updated Edge Functions

   supabase functions deploy ingest-thought --no-verify-jwt
   supabase functions deploy open-brain-mcp --no-verify-jwt

5. Set your environment variables

   export INGEST_URL="https://YOUR_PROJECT_REF.supabase.co/functions/v1/ingest-thought"
   export INGEST_KEY="your-ingest-key"

   Add to ~/.zshrc to make them permanent.

6. First dry run

   deno run --allow-net --allow-read --allow-write --allow-env \
     scripts/pull-gmail.ts --dry-run --window=24h

   Authorize via the URL it prints. Check the output — are long emails showing [N chunks]? Is the content preview clean? When it looks right, drop --dry-run to go live.

7. Scale up

   deno run --allow-net --allow-read --allow-write --allow-env \
     scripts/pull-gmail.ts --window=30d --labels=SENT,STARRED

   Re-running is safe — the sync log tracks ingested IDs and skips duplicates.

Label Strategy

Pro tip: Gmail labels become searchable metadata. The script stores all labels on each ingested thought, so you can later ask your AI "show me everything I tagged as Project X."

Database Changes

These are nullable — your existing thoughts are unaffected. The migration uses IF NOT EXISTS, safe to run on a live database.

Updated Edge Functions

ingest-thought now accepts parent_id, chunk_index, and extra_metadata. Has 3-attempt retry logic for OpenRouter failures. Caps embedding input at 8,000 characters. Crashes on startup if INGEST_KEY is unset (rather than silently accepting all requests).

open-brain-mcp search now fetches 3× the requested limit and deduplicates chunks from the same parent — if 3 chunks of a long email match your query, you get one result with a note. New tool: email_sync_status.

New Metadata Fields

{
  "type": "observation",
  "topics": ["strategy", "product"],
  "people": ["Alice", "Bob"],
  "sentiment": "positive",
  "source": "gmail",
  "gmail_labels": ["SENT", "IMPORTANT"],
  "gmail_id": "18e4f2a...",
  "gmail_thread_id": "18e4f1..."
}

What's Next

The partial protection: The ingest-thought function sends content to OpenRouter to extract structured JSON metadata — not to reason freely. The structure acts as a summarization barrier. The embedding step is purely mathematical. Neither step is a high-risk surface.

Where the real risk lives: When you ask your AI to "summarize everything I emailed about Project X" — that's when retrieved email content enters the AI's context window alongside your instructions.

Practical mitigations:

• Stick to SENT as your primary label — you wrote those emails
• Be more cautious with STARRED or INBOX (content from untrusted senders)
• The gmail.readonly OAuth scope means the script can never send email on your behalf
• If an AI client behaves strangely after a memory search, check what was retrieved

Right now the script is manual. Here's where things stand and where they're going.

Option 1: cron (works today)

# Add to crontab: crontab -e
# Runs every Monday at 8am
0 8 * * 1 cd /path/to/monkeyrun-open-brain && \
  INGEST_URL="..." INGEST_KEY="..." \
  deno run --allow-net --allow-read --allow-write --allow-env \
  scripts/pull-gmail.ts --window=7d --labels=SENT,STARRED

Option 2: OpenClaw (the right long-term home)

OpenClaw is built for exactly this — scheduled tasks, local script execution, Telegram notifications when done. If you're already running it, give it this prompt:

"Add a weekly cron job that runs every Monday at 8am. It should cd into my Open Brain project directory and run deno run --allow-net --allow-read --allow-write --allow-env scripts/pull-gmail.ts --window=7d --labels=SENT,STARRED. When it finishes, send me a Telegram message with the summary line from the output."

OpenClaw builds the cron job, hooks it into its scheduler, and you get a weekly brain sync with a Telegram confirmation. The Gmail OAuth token lives in the project directory — no extra credential setup needed.

Option 3: MCP trigger (roadmap)

The cleanest eventual solution: a pull_emails MCP tool that lets you trigger a sync from any AI client just by asking for one. Deferred because it requires moving the OAuth token server-side or building a webhook architecture. On the roadmap as the system matures.

Script Options

Troubleshooting