Build from evidence,
not from mythology.

The core package requires Python 3.10 or newer. WhatsApp evidence stays on the machine unless you deliberately connect another provider.

git clone https://github.com/Amal-David/whatsapp-llm
cd whatsapp-llm
pip install -e .

The multi-chat workflow separates owner evidence, relationship context, and owner self-report. It produces digital_self.v1, a portable profile with hashes and aggregates rather than owner message bodies.

1. List sources

WhatsApp for Mac is opened through SQLite read-only mode. Listing chats does not print message bodies.

living-brain self chats --source whatsapp-mac

2. Interview the owner

living-brain self interview \
  --owner-name "Your Name" \
  --output ./private/self-interview.yaml

3. Build and validate

living-brain self build \
  --source whatsapp-mac \
  --chat "first-chat-id" \
  --chat "second-chat-id" \
  --owner-name "Your Name" \
  --interview ./private/self-interview.yaml \
  --output ./private/digital-self.json

living-brain self validate ./private/digital-self.json

Use --all-chats only after reviewing the source list. For a local wacli mirror, pass --source wacli --path ~/.wacli/wacli.db.

Mini-Me asks a narrower question: how does one selected person tend to reply in this conversational setting? It starts with an authorized WhatsApp text export.

living-brain workbench --port 7861

1. 01
   Upload

   Choose an export and select the participant whose messages may be used.

2. 02
   Consent

   Confirm authority before the dataset is generated.

3. 03
   Separate

   Third-party messages are withheld by default; target replies are redacted.

4. 04
   Hold out

   Whole conversations, not adjacent messages, define train and evaluation splits.

The archive includes canonical examples, style summaries, training-ready rows, preference pairs, and held-out evaluation rows. Keep the result relationship-scoped unless broader evidence supports generalization.

Run a deterministic synthetic walkthrough. It reads no WhatsApp data and calls no external model.

living-brain brain guide \
  --demo \
  --workspace ./private/digital-self-demo \
  --as-of 2026-07-11T12:00:00+00:00

The typed-state evaluator does not collapse human representation into one flattering number.

• Partial evidenceChats show selected behavior, not a complete person or hidden motives.
• Private sourcesKeep chats, keys, interviews, profiles, evaluation rows, and state snapshots out of source control.
• Version-sensitive accessWhatsApp for Mac uses a private schema; wacli is not an official API.
• Incomplete deletionComplete propagation through every derived artifact is not implemented yet.
• No authorityGenerated output cannot send, promise, transact, or represent the owner.

pip install -e ".[dev]"
PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 pytest tests -q -o addopts=''
python -m compileall -q living_brain tests

CI repeats tests, compilation, CLI smoke checks, package builds, Ruff, and mypy on Python 3.10, 3.11, and 3.12.