A zero-dependency SQLite memory layer for AI agents that learns what works. Most memory tools store what you tell them; metabrain closes the loop—a pattern recorded enough times graduates into a hypothesis, every outcome becomes an experiment for or against it, and proven hypotheses graduate into preferences the agent runs on.

Deterministic by construction: using the API correctly populates every table as a side effect. Seven tables, one file, stdlib sqlite—no vector database, no server, no API key. General across self-learning content engines, lead capture, and self-improving job applications.

Self-evolving Claude Code plugin built from failure paths, not theory. Every pattern earned by breaking something first, then encoding the fix so it never breaks the same way twice.

AgentDB-first methodology where every artifact reads on start, writes on end. Multi-agent orchestration with contracts, checkpoints, and verdicts. An experiment engine that proves which workflows actually hold up. Active development, open source.

Practical benchmark for local and open-source LLMs—21 programmatically-verified tests that mirror real work: extracting structured data, finding bugs, resisting prompt injection, reasoning through dependency chains. Not another MMLU wrapper.

Every test has a deterministic verifier—no LLM-as-judge, no vibes. Benchmarks any local model against Claude haiku/sonnet/opus tiers through headless Claude Code, no API key required. Collaborative: run it on your hardware, submit results, build a shared performance map.

A2A communication protocol with 3x semantic density. Natural language is inefficient for agent coordination—ambiguity, wasted tokens, latency.

A self-evolving agent framework—AI agents that get better every time they run. Normally you deploy an agent, then burn weeks hand-tuning prompts, swapping models, and adjusting temperatures, and redo it all whenever user behavior shifts. Armature wraps the agent in a reinforcement-learning loop so it improves automatically from its own outcomes.

Reinforcement learning + evolutionary algorithms + self-improving policies: weak behaviors die off, strong ones survive and propagate. Published to PyPI as armature-ai and integrated into HeyContext.