AgentOS

AgentOS gives TypeScript agents a memory that persists, a toolkit that can grow mid-session, and one dispatch interface across eleven model providers. Three things accumulate as a session runs: memory, the tool surface and an optional personality vector, each configurable and observable rather than buried in a prompt.

Cognitive memory sits at the center, built from eight mechanisms out of memory research, among them Ebbinghaus decay, retrieval-induced forgetting and reconsolidation. On LongMemEval-S it reports 85.6 percent with a gpt-4o reader, and 70.2 percent on LongMemEval-M, with confidence intervals and single-command reproduction.

The rest of the runtime is parts you switch on as needed.

• Tool forging: an agent writes a TypeScript function with a Zod schema, an LLM judge approves it, and it runs sandboxed before joining the catalog.
• HEXACO personality: off by default, and when on it weights retrieval, routing and tool selection from the kernel, not the prompt.
• Soul files: identity, voice and hard limits live in a markdown workspace whose memory wiki is the source of truth.
• Orchestration: workflow DAGs, agent graphs with cycles, goal-driven missions and six coordination strategies.
• Retrieval: seven vector backends, GraphRAG, HyDE and Cohere reranking.
• Guardrails: five security tiers with packs for PII, code safety and content policy, plus 37 platform adapters.

Install is one npm package, and the provider is detected from the environment when none is named. Sessions keep a lossless transcript independent of memory, bounded by default and replaceable by reseeding. Prompt caching works on every provider without setup, under an Apache 2.0 license.