What is alaya?

Alaya is a local-first system for teams or individuals who want AI assistance in making product or operational decisions, but need a way to keep that AI accountable. The README is written in Chinese, but the core idea is that every cycle of work goes through a fixed loop: make a prediction, take an action, observe what happened, figure out what went wrong, distill that into a knowledge entry, and use that entry in the next cycle. The goal is not unlimited AI automation but AI operating under strict, auditable constraints. The system runs entirely on your own machine using a local SQLite database. It includes five AI agents that each play a specific role: one selects goals and references prior knowledge, one collects feedback and external signals, one forms task packages, one distills observations into knowledge, and one manages the knowledge library including merging duplicates and expiring outdated entries. By default these agents use a built-in mock model so no external AI service is required, but you can connect any OpenAI-compatible endpoint to use a real language model. A central feature is the Human Gate system. Before the AI can act on high-risk decisions, a human must approve through either the web interface or a Telegram notification on their phone. There are three gate types: one for direction decisions, one for deciding whether a signal is worth adding to the knowledge base, and one for risk or cost decisions. The Telegram integration lets you receive alerts and approve or reject certain gates from your phone without needing to stay at a computer. The knowledge base is designed to be more structured than plain notes. Each entry tracks confidence level, status, source, references, expiry, and whether it has been superseded by a newer entry. The system applies time decay to entries, lowering their confidence over time if they have not been recently verified, and excludes stale or conflicting entries from influencing new decisions. The project ships as two packages: a pure TypeScript core that contains only the flywheel logic and runs without a database or web layer, and a full web application that adds Express, React, SQLite, a scheduler, and the Telegram integration. You start it with Node.js and access it at a local URL. A demo project is seeded on first run to show four complete cycles of the flywheel with predictions, observations, and human gates included.