AgentOS

Key features

• Cognitive memory with forgetting curves Eight neuroscience-backed mechanisms, including Ebbinghaus decay, reconsolidation, and retrieval-induced forgetting, keep long-term context while letting irrelevant detail fade.[1]

• Runtime tool forging When no tool covers a sub-task, the agent writes a TypeScript function with a Zod schema, a separate language-model judge approves it, and it runs in a hardened sandbox before joining the session catalog; proven tools can export as SKILL.md skills.[1]

• Multi-agent teams in six lines Six coordination strategies with shared memory, inter-agent messaging, and human-in-the-loop gates, plus workflow DAGs, agent graphs, and goal-driven missions with checkpointing.[1]

• Eleven providers with fallback chains OpenAI, Anthropic, Gemini, Groq, Ollama, OpenRouter, Together, Mistral, xAI, and CLI-based options route with automatic failover and key rotation, alongside zero-config prompt caching on every provider.[1][2]

• Retrieval across seven vector backends Postgres with pgvector, Qdrant, Neo4j, and others combine with GraphRAG, HyDE, and Cohere reranking, while a capability-discovery engine replaces static tool dumps with semantic search.[1]

• Five guardrail tiers Dangerous, permissive, balanced, strict, and paranoid tiers pair with PII redaction, prompt-injection detection, code safety, grounding checks, topicality enforcement, and content filtering; guardrails can even override approvals on destructive actions.[1]

• Sessions that survive long runs Each session keeps tool calls, results, and model reasoning up to a roughly 120K-token cap, with reseed controls and an optional HEXACO personality vector that biases retrieval and routing without changing runtime behavior.[1]

• Thirty-seven channels plus voice and observability Telegram, Discord, Slack, WhatsApp, and webchat adapters sit beside ElevenLabs, Deepgram, and telephony integrations, with OpenTelemetry tracing, a usage ledger, cost guards, and circuit breakers.[1]

Who will love it

• Memory, sessions, and orchestration arrive in one npm package, so a Node.js application gains agents without adopting a Python stack. Installation on Node.js 18 or later with SQLite by default means a prototype starts without provisioning managed services.[1][2]

• PII redaction, tiered guardrails, approval gates, and fully offline operation with Ollama, SQLite, and in-memory vectors fit teams that must keep data on their own infrastructure. The project documents data minimization, consent, and deletion controls for GDPR-aligned self-hosting.[1][2]

• Six team strategies, seven vector backends, and prompt caching cover ambitious agent systems where LangChain-style assembly would mean wiring many pieces by hand. Capability discovery replaces static tool dumps with semantic search, which keeps large tool catalogs affordable in tokens.[1]

Who should avoid it

• The LangChain, LangGraph, and CrewAI ecosystems match a Python codebase better; adopting AgentOS means switching languages for the agent layer.[1]

• The SDK needs TypeScript, and the friendlier Wunderland CLI with its setup wizard is explicitly a preview whose surface may still change.[1]

• Roughly a year old with a few hundred stars at inspection time, AgentOS offers published benchmarks but little independent confirmation of production behavior. Its rapid release cadence shows velocity, not maturity, so risk-averse buyers should weigh that accordingly.[1][2]

Alternatives and compatibility