A chatbot that refuses to make things up about you

What it does

Agent-Me is an open-source FastAPI + React system for building a personal AI agent from your own knowledge. You feed it version-controlled Markdown — projects, preferences, decisions — and it answers questions through a sequential pipeline: Planner, Researcher, Critic, Writer, and an optional Verifier. The public repo ships with fictional example data; your real identity lives in a separate private workspace.

The interesting bit

The Critic is an evidence-sufficiency gate: when retrieved knowledge doesn’t support an answer, it blocks synthesis and the Writer returns a fixed “insufficient evidence” response instead of improvising. Every run exposes a public trace — role outcomes, evidence excerpts, metrics — so you can see exactly why the twin said what it said.

Key highlights

• Memory substrate is plain, reviewable, version-controlled Markdown — no opaque vector store as the source of truth.
• Deterministic local retrieval and evaluation; the default setup needs no API key and runs fully offline.
• Typed Python contracts (Plan, EvidenceBundle, Critique, WrittenAnswer, Verification) between agent roles.
• Versioned evaluation fixtures cover supported, unsupported, adversarial, and boundary requests, run in CI.
• Optional OpenAI-compatible provider mode (e.g. OrcaRouter) for the simple chat path only; the collaboration pipeline stays local.

Caveats

• The Verifier checks citation paths and output invariants — it does not prove factual truth or semantic entailment (the docs are refreshingly honest about this).
• Automatic temporal reasoning, knowledge graphs, and richer identity models are explicitly future work, not current features.
• Only POST /api/v1/chat crosses the provider boundary; if you use a hosted LLM there, your question and retrieved context go with it.

Verdict

Worth a look if you’re interested in agent architectures where evidence and inspectability are first-class, or you genuinely want a self-hosted personal knowledge assistant. Skip it if you expected a polished product — this is a well-documented exploration with example data, and it knows it.