MeantByMe

Completed with AI. Meant by me.

简体中文 · Project Hub · Web Demo · Documentation

Overview

MeantByMe is a consent-first communication agent for people whose intent remains relatively clear while speech is fragmented, slow, impaired, or temporarily unavailable.

The system may use speech recognition, context, and personal memory to propose candidate expressions, but these inputs are evidence—not authority. AI may help complete a sentence; the person must explicitly confirm its meaning, verified-memory write, and personal-voice authorization.

Core principle: AI proposes; the patient decides. Models produce evidence; the deterministic runtime controls consent.

AdventureX Hackathon 2026

MeantByMe was created as an AdventureX Hackathon 2026 project. HumanLoop Lab selected one of the event's five Topic Tracks, Reverse, and entered three additional Sponsor Tracks.

Category	Track	Result
Topic Track	Reverse	Runner-up
Sponsor Track	Qoder	Third Prize
Sponsor Track	Superun.ai	Second Prize
Sponsor Track	viaim / iFLYTEK	Excellence Award

Co-evolution framework

“Co-evolution” means auditable personalization based on explicit, provenance-preserving confirmation. It does not allow AI to learn, authorize, or write Gold Memory from silence, timeout, caregiver action, or model confidence. This is a product-concept diagram, not a clinical or security certification.

Safety contract

• Unconfirmed candidates never use the patient's personal voice.
• Silence, timeout, caregiver action, confidence, and device presence are not consent.
• AI output never writes directly to Gold Patient Memory.
• Caregiver context remains distinguishable from patient-confirmed intent.
• Rejected candidates never become patient preferences.
• Medical, legal, financial, and major relationship expressions require stricter confirmation.
• Stop, Back, None of these, and Switch input method remain available.
• Every external expression is traceable to evidence, AI completion, confirmation method, and authorization state.

The complete engineering invariants are in AGENTS.md; accepted implementation decisions are in DECISIONS.md.

Simple flow

fragmented speech or alternative input
→ ASR and contextual evidence
→ patient-scoped verified-memory retrieval
→ clarification or candidate expressions
→ explicit patient selection
→ private final readback
→ current-expression authorization
→ personal-voice output
→ Expression Receipt
→ idempotent verified-memory update

Personal voice is used only when both the final content and the current expression authorization have been explicitly confirmed.

Repository status

This repository contains several implementation tracks. Branch names describe development history rather than clean component boundaries.

Track	Branch	Current contents	Verified tests
Canonical baseline and project hub	main	Deterministic mock runtime, schemas, safety policies, documentation, GitHub Pages	24
Integrated Web/backend prototype	develop, frontend	Same commit: runtime, gateway, responsive Web BFF/demo, cloud adapters, evaluation	142
Headset/mobile experiment	feature/earPhones	iOS, Viaim headset adapter, command/QA and profile-storage experiments	172

Test counts were independently verified on 2026-07-26 with Python 3.11.8. develop and frontend currently point to the same commit. The mobile branch is a broad iOS, backend, database, and memory experiment—not a merge-ready mobile-only patch.

See the branch audit and current status for exact snapshots and integration risks.

Run the canonical baseline

Requires Python 3.11:

python3.11 -m venv .venv
./.venv/bin/python -m pip install -e '.[test]'
./.venv/bin/python -m meantbyme --mode mock
./.venv/bin/python -m pytest

The mock flow uses fixture ASR, patient-scoped SQLite memory, deterministic candidates, cached TTS, and no network. It exits successfully only after the session reaches completed.

For the Web/gateway prototype or iOS experiment, switch to the corresponding branch and follow its component README. Never place provider secrets, real patient data, personal voice material, or production databases in source, frontend bundles, logs, fixtures, or commits.

Documentation and project boundary

• Documentation index and authority
• Technical architecture
• Agent runtime
• Memory and personalization
• Security, privacy, and consent
• Integration plan
• Evaluation and testing
• Contribution guide
• Security policy

MeantByMe is a research and communication-assistance prototype. It is not a medical diagnosis system, treatment decision system, mind-reading system, or clinically certified medical device, and it should not be the sole channel for emergency communication.

Reference architecture

This is the target/reference architecture across implementation tracks, not a claim that every displayed provider is deployed on main. Probabilistic services produce transcripts, candidates, and audio; only the deterministic control plane may advance confirmation, authorize personal voice, create an Expression Receipt, or write verified memory. All storage and retrieval remain patient scoped.

License

Copyright © 2026 MeantByMe contributors. Released under the MIT License. The license permits use of the software; it does not grant rights to patient data, voice recordings, voice identities, third-party models, vendor SDKs, names, or trademarks.