What is AgentOS?

AgentOS is the open-source TypeScript runtime for autonomous AI agents. Each agent is a Generalized Mind Instance (GMI) with a HEXACO personality, eight neuroscience-backed memory mechanisms, runtime tool forging inside a V8 sandbox, and a classifier-driven memory pipeline that decides per-query whether to retrieve, which architecture to use, and which reader to dispatch.

AgentOS powers production agents on Wilds.ai, Wunderland.sh, and Paracosm. The package ships under the Apache-2.0 license with 3,866+ tests in CI, full TypeScript types, and zero vendor lock-in.

npm install @framers/agentos

Cognitive Architecture, Not Just an Orchestrator

Most agent libraries are routing layers — they accept a prompt, pick a tool, return an answer. AgentOS treats the agent as a mind with persistent identity, memory, and behavioral adaptation across sessions.

HEXACO Personality

Every GMI carries a HEXACO-60 personality vector: Honesty-Humility, Emotionality, eXtraversion, Agreeableness, Conscientiousness, Openness. The six traits modulate retrieval bias (a high-Openness agent surfaces more counterintuitive memories), response style (a high-Emotionality agent leans into sentiment), and tool selection (a high-Conscientiousness agent picks higher-fidelity tools). Personality is not a system prompt template — it shapes which memories get encoded, how they decay, and what the agent actually picks from the retrieval candidates.

Eight-Mechanism Cognitive Memory

The memory engine implements eight neuroscience-backed processes: reconsolidation (memories rewrite when re-accessed), retrieval-induced forgetting (RIF), involuntary recall (proactive surfacing), feeling-of-knowing (FOK metacognition), gist extraction (compressed summaries), schema encoding (typed slots), source decay (provenance fades on the Ebbinghaus forgetting curve), and emotion regulation (mood-tagged retrieval). The result is an agent that remembers what mattered, forgets what didn't, and can correct itself when memory and reality diverge.

Classifier-Driven Memory Pipeline

The standout feature: a three-stage LLM-as-judge classifier pipeline that gates retrieval per query. Stage 1 decides whether memory is even needed. Stage 2 picks the retrieval architecture (canonical-hybrid, observational-memory-v10, observational-memory-v11). Stage 3 picks the reader tier (gpt-4o for hard categories, gpt-5-mini for easy ones). Stages 2 and 3 reuse the Stage 1 classification so the full pipeline costs one classifier call per query, not three.

Internal evaluation on LongMemEval-S Phase B at N=500 with 10k-resample bootstrap confidence intervals: 85.6% accuracy [82.4%, 88.6%] at $0.0090 per correct answer. The same harness measures Mastra OM gpt-4o at 84.2% (matches their published number) and re-runs EmergenceMem Simple Fast at 80.6% (apples-to-apples in our harness; their published headline is 79.0%). AgentOS lands +5.0 percentage points over EmergenceMem at 6.5× lower cost-per-correct. Full reproducible run JSONs and methodology in the agentos-bench repo.

Sixteen LLM Providers, Automatic Fallback

AgentOS ships native adapters for OpenAI, Anthropic, Gemini, Groq, Ollama, OpenRouter, Together, Mistral, xAI, Claude CLI, Gemini CLI, plus five image/video providers. When the primary returns HTTP 402/429/5xx, network fails, or auth breaks, generateText walks a canonical fallback chain using whichever API keys are present. No extra imports, no chain construction. Strict single-provider routing is one parameter away for billing isolation, capability auditing, or provider-pinned tests.

Seven Vector Backends, Four Retrieval Strategies, GraphRAG

The RAG pipeline supports InMemory, SQL, HNSW, Qdrant, Neo4j, pgvector, and Pinecone. Four retrieval strategies (vanilla similarity, hybrid BM25 + dense, query rewriting, GraphRAG) cover the full spectrum from "small dataset on a laptop" to "production knowledge graph with citation paths."

Multi-Agent Teams With Six Strategies

Six coordination strategies for multi-agent work: sequential, parallel, debate, review-loop, hierarchical, graph. Each agent has its own HEXACO profile, memory, and skill set; teams share memory through scoped channels with HITL approval gates. The orchestration runtime offers three authoring surfaces — workflow() for deterministic DAGs, AgentGraph for cycles and subgraphs, mission() for goal-driven adaptive planning — all compiling to the same graph runtime with persistent checkpointing.

Five-Tier Guardrails

Security ships built in. Five tiers from dangerous (no guardrails) to paranoid (full pipeline + circuit breakers). Six guardrail packs cover PII redaction, ML-based classifiers, topicality enforcement, code-execution safety, grounding-against-context, and content policy. Prompt injection patterns are caught at the pre-LLM stage; outputs are signed with HMAC-SHA256 to maintain an intent-chain audit trail.

How AgentOS Compares

The core differentiation is the cognitive layer — personality + memory + classifier-driven retrieval — that other frameworks treat as out-of-scope.

Quick Start — A Personality-Driven Agent in 10 Lines

import { agent } from '@framers/agentos';

const research = agent({
  name: 'researcher',
  personality: { O: 0.9, C: 0.85 }, // high openness, high conscientiousness
  memory: { backend: 'pgvector', mechanisms: 'all' },
  guardrails: { tier: 'balanced' },
  providers: ['openai', 'anthropic'], // automatic fallback
});

const answer = await research.ask('Find the most-cited 2026 paper on cognitive AI agents.');

The agent has persistent memory across calls, falls back from OpenAI to Anthropic on errors, blocks prompt injection at the guardrail stage, and shapes retrieval by its high-Openness profile.

The AgentOS Family on npm

Production Deployments

• Wilds.ai — every NPC, narrator, and game master runs on AgentOS. HEXACO companions with mood drift, voice synthesis, and 12-genre game generation.
• Wunderland.sh — open-source OpenClaw fork distributing AgentOS as a personal AI assistant across 37 messaging channels.
• Paracosm — counterfactual world simulation; AgentOS provides the GMI substrate, HEXACO actors, and runtime tool forging.

Frequently Asked Questions

What is AgentOS?

AgentOS is an open-source TypeScript runtime for autonomous AI agents with cognitive memory, HEXACO personality, multi-agent orchestration, and a classifier-driven retrieval pipeline. It is published on npm as @framers/agentos under the Apache-2.0 license.

Is AgentOS free and open-source?

Yes. AgentOS is Apache-2.0 licensed with the source on GitHub. You can use it commercially without royalty, modify it, and self-host it. There are no paid tiers of the runtime itself — paid services like Wilds.ai are built on top of it.

How does AgentOS differ from LangGraph?

LangGraph is a graph orchestration library — it routes prompts and tools through nodes. AgentOS is a cognitive runtime: agents have persistent HEXACO personalities, eight-mechanism memory with reconsolidation and forgetting curves, runtime tool forging, and a classifier pipeline that decides per query whether memory is even needed. Orchestration is one component of AgentOS, not the whole framework.

What LLM providers does AgentOS support?

Sixteen native providers: OpenAI, Anthropic, Gemini, Groq, Ollama, OpenRouter, Together, Mistral, xAI, Claude CLI, Gemini CLI, plus five image and video providers. Automatic fallback chains kick in on retryable errors (HTTP 402/429/5xx, network failures). You can also opt out for strict single-provider routing.

Does AgentOS work with voice?

Yes. The voice pipeline supports ElevenLabs, Deepgram, and OpenAI Whisper for STT/TTS, plus telephony adapters for Twilio, Telnyx, and Plivo. Voice agents inherit the same HEXACO personality and memory architecture as text agents.

How does AgentOS handle agent memory?

Eight neuroscience-backed mechanisms run on top of seven supported vector backends. Memory is encoded with HEXACO-modulated salience, consolidates during sleep cycles, decays per the Ebbinghaus forgetting curve, and gets reconsolidated when re-accessed. A three-stage classifier pipeline gates retrieval per query — most queries skip retrieval entirely, the rest get the right architecture and reader for the question type.

Can I use AgentOS without Wilds or Wunderland?

Yes. AgentOS is the underlying runtime; Wilds.ai and Wunderland.sh are products built on it. You can install @framers/agentos directly and build your own agents, teams, or runtimes. The package is provider-agnostic and self-hostable.

How do I get started with AgentOS?

Run npm install @framers/agentos, set an API key for any supported provider, and call agent() or generateText(). The full quickstart is at docs.agentos.sh. The example repository is at github.com/framersai/agentos — clone it for runnable code samples covering memory, multi-agent teams, RAG, and orchestration.

Get Started

AgentOS is the cognitive substrate behind every AI agent we ship. Install it, fork it, build on top of it.

Read the docs at agentos.sh → | Browse the source on GitHub → | Install from npm → | Join the Discord →