Show HN: ANM – a local multi-agent reasoning system with Web-of-Thought

1 point

5 months ago

I built ANM to explore whether multi agent reasoning can reduce hallucinations under tight hardware constraints.

ANM replaces single model Chain of Thought with Web of Thought. Multiple domain specialists collaborate, cross check each other, and are gated by a verifier and a runtime Law Book.

Key ideas include:

Web of Thought orchestration across 12 specialists

PAST ONLY memory that stores observations rather than truths

Metacognitive confidence calibration

Runtime constitutional governance without retraining

Fully local execution on consumer hardware, tested on a MacBook Air M2 with 16GB RAM

This is a research prototype, not a product. Latency is high, models are small, between 1.5B and 3B parameters, and results are qualitative rather than competitive benchmarks.

The paper, code, and benchmark traces are fully open source. I would really appreciate critical feedback from people working on reasoning, agents, or reliability.