AgentOS
AgentOS is an open-source TypeScript framework for AI agents with persistent cognitive memory, runtime tool forging, multi-agent orchestration and one dispatch interface across 11 LLM providers.
AgentOS is a TypeScript framework for building AI agents that remember across sessions, adapt, and can write their own tools. Its runtime tool forging lets an agent write a TypeScript function with a Zod schema, have an LLM judge approve it, and run it in a hardened node:vm sandbox before it joins the agent's tool catalog for the rest of the session.
Its persistent cognitive memory implements eight neuroscience-inspired mechanisms, including Ebbinghaus decay, retrieval-induced forgetting, reconsolidation and source-confidence decay, and reports 85.6% accuracy on the LongMemEval-S memory benchmark and 70.2% on LongMemEval-M. It supports an optional HEXACO personality model, six multi-agent orchestration strategies, guardrails, and voice, with more than 100 extensions and 88 skills that auto-load at startup.
Sessions carry a lossless conversation transcript independent of the memory subsystem, bounded by default eviction past roughly 120,000 tokens, with support for atomic history replacement, checkpointing and per-send generation overrides. It installs with npm and is licensed under Apache-2.0.