framerslab/agentos

AgentOS – TypeScript Framework for AI Agents

What it is

• An open‑source TypeScript library that lets you build LLM‑powered agents that can remember past interactions, adopt a personality, and even create new tools while they run.
• It works with 13 different LLM providers (OpenAI, Anthropic, Gemini, Ollama, etc.) and ships a large ecosystem of extensions, skills and channel adapters.

Key capabilities

Feature	What it does
Cognitive memory	Stores conversation turns in a vector store, applies eight neuroscience‑inspired mechanisms (e.g., reconsolidation, retrieval‑induced forgetting) and decays memories over time.
Runtime tool forging	When a task needs a tool the agent doesn’t have, it writes a JavaScript function with JSON‑Schema I/O, runs supplied test cases, gets an LLM judge to approve it, and then adds the tool to the session’s toolbox. Execution can be sandboxed in node:vm or QuickJS WebAssembly.
HEXACO personality (optional)	A six‑trait personality vector (openness, conscientiousness, etc.) is injected into prompts and influences memory encoding and routing decisions.
Generalized Mind Instances (GMIs)	Persistent “full‑runtime” agents that keep a persona, mood, reasoning trace and sentiment‑driven metaprompts across sessions.
Multi‑agent orchestration	agency() lets you compose several agents with six coordination strategies (sequential, parallel, debate, review‑loop, hierarchical, graph).
RAG pipeline	Built‑in support for seven vector backends, four retrieval strategies, GraphRAG, HyDE and reranking.
Guardrails & safety	Packs for PII filtering, code safety, grounding, etc., plus a circuit‑breaker and cost‑guard.
Voice & telephony	Integration with ElevenLabs, Deepgram, Whisper and telephony providers (Twilio, Telnyx, Plivo).
Observability	OpenTelemetry, usage ledger, cost monitoring.

Typical usage

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

const tutor = agent({
  provider: 'anthropic',          // defaults to Claude‑sonnet‑4‑6
  instructions: 'You are a patient CS tutor.',
  personality: { openness: 0.9, conscientiousness: 0.95 },
});

const session = tutor.session('student-1');
await session.send('Explain recursion with an analogy.');
await session.send('Can you expand on that?'); // remembers prior turn

The session records every tool call, its results and the model’s “thinking”, and can replay the whole history later.

Why it stands out

• Claims the best open‑source memory benchmark scores (85.6 % on LongMemEval‑S, 70.2 % on LongMemEval‑M) with a transparent methodology.
• Introduces runtime tool forging—agents can extend their own toolbox without developer intervention.
• Provides an optional personality model and a full‑runtime GMI that adds mood‑aware memory and self‑improvement tools.
• A rich plugin ecosystem (100+ extensions, 88 pre‑built skills) and a CLI/daemon (wunderland) for production deployment.

Installation & docs

• Install via npm: npm install @framers/agentos.
• Full docs, quick‑start, API reference and benchmark details are hosted at docs.agentos.sh.
• Community chat on Discord and a public benchmark repo (agentos-bench).

Bottom line AgentOS is a comprehensive, TypeScript‑first framework for building sophisticated, memory‑rich AI agents that can adapt, create tools, and operate in multi‑agent teams, all while supporting a wide range of LLM back‑ends and safety features.