Digital Self

An experimental, local-first toolkit for building inspectable representations of one person from authorized WhatsApp conversations.

The project explores two scopes:

Approach	Input	Result
Mini-Me	One selected person's messages, usually from one chat	A relationship-scoped writing-style dataset with held-out evaluation
Digital Self	Owner messages across several selected chats, plus an optional interview	A versioned profile with evidence, time, uncertainty, relationships, and optional typed state

Neither approach reconstructs a person or a mind. The output is an experimental model of selected evidence that the owner must inspect, correct, and judge.

The installed command remains living-brain for compatibility.

Current State

The repository currently supports:

• read-only chat discovery and ingestion from WhatsApp for Mac
• ingestion from a local third-party wacli linked-device mirror
• explicit chat selection, stable pseudonyms, and owner interviews
• digital_self.v1 profiles with evidence hashes, provenance, confidence, temporal validity, contradictions, and relationship-specific style
• private held-out evaluation rows and separate text-free summaries
• migration to typed state with coverage, correction, scoped simulation, and independent evaluation axes
• a deterministic synthetic walkthrough that writes 12 auditable artifacts
• a narrow Mini-Me workbench for consented, single-person style datasets

Install

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

Build A Multi-Chat Digital Self

List available chats without printing message bodies:

living-brain self chats --source whatsapp-mac

Create a private owner interview:

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

Build and validate a profile from explicitly selected chats:

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. To use a wacli mirror, pass --source wacli --path ~/.wacli/wacli.db. The self evaluate command creates private held-out rows and a separate text-free summary.

Build A Mini-Me

Mini-Me is the narrower workflow for learning how one selected person tends to write in a particular conversational setting.

living-brain workbench --port 7861

Upload an authorized WhatsApp text export, select the participant, confirm consent, and review the generated archive. The workbench produces canonical examples, style summaries, training-ready rows, preference pairs, and held-out evaluation rows. Third-party messages are withheld by default.

Treat the result as relationship-scoped unless evidence from other chats shows that the same behavior generalizes.

Inspect The Typed-State Workflow

Run the complete deterministic walkthrough with synthetic data only:

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

The walkthrough performs source selection, initial state creation, coverage analysis, adaptive interview, versioning, inspection, owner correction, simulation, and evaluation. It reads no WhatsApp data and calls no external model.

The 12 generated files are workflow stages and receipts, not 12 models or a single replica score. Their exact purpose is documented in the Digital Self explainer.

Evidence And Evaluation

The owner selects local sources. The system normalizes messages, separates owner evidence from third-party context, and records provenance. Observed behavior remains a candidate rather than a fact; claims and style stay scoped by time and relationship.

Owner message bodies are not copied into the portable digital_self.v1 profile. The profile stores hashes, aggregates, provenance, and relationship-safe metadata. Private evaluation rows may contain conversation text and are written with owner-only permissions; their summary contains counts and coverage only.

Mini-Me checks held-out replies for style, context, privacy, invention, and memorization. The typed-state evaluator separately checks behavior, time, relationships, attribution, decisions, calibration, privacy, authority, and explicit owner judgment. Privacy failures block a pass. There is no single "percent cloned" score.

Limits

• WhatsApp conversations are partial, relationship-specific evidence, and the Mac app uses a private schema that may change.
• wacli is a third-party linked-device mirror, not an official WhatsApp API or a guaranteed complete archive.
• Automatic raw-message-to-typed-state extraction and complete deletion propagation are not finished.
• The synthetic walkthrough proves software contracts, not real-person fidelity.
• The project has no authority to send, promise, transact, or represent the owner.

Keep raw chats, pseudonym keys, interviews, profiles, evaluation rows, and typed state snapshots private and out of source control.

Documentation

• Digital Self explainer: extraction, conversion, Mini-Me artifacts, evaluation, and the 12 guided-run files
• Typed Digital Self PoC: schemas, migration, correction, simulation, and verification
• Research synthesis: evidence, constraints, architecture, and non-goals
• Research council README: corpus structure and reproducibility workflow

Development

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

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

License

MIT. See LICENSE.