An open-source framework that treats AI agents not as disposable “tools” but as team members who keep their memories. Under the banner of “Organization-as-Code,” each agent (an Anima) is given a name, role, personality, memory, and schedule, and works autonomously within a hierarchy of supervisors and subordinates. It is built by xuiltul, a psychiatrist and engineer, and its defining feature is a memory system inspired by brain science ─ agents remember events, consolidate them into knowledge overnight, and gradually forget what goes unused. Released in February 2026, it is free to self-host under the Apache License 2.0.
Key Features
- Agent teams with a real org structure: Setting a
supervisorfield on each Anima is all it takes to build a hierarchy. Manager roles automatically receive supervisor tools ─ task delegation, progress tracking, restarting subordinates ─ so a human only needs to talk to the leader - Brain-inspired memory system: When a message arrives, related memories (sender profile, recent activity, important knowledge, past episodes) are recalled automatically (priming), and an agent can also search on its own with
search_memory. At night, the day’s episodes are consolidated into knowledge, while rarely used memories are progressively forgotten and archived. Nothing is crammed into the context window - Autonomous operation that does not stop: A heartbeat periodically reviews the situation and decides what to do next, per-Anima cron jobs handle daily reports, weekly digests, and monitoring, and a task engine resolves dependencies and runs independent tasks in parallel. Agents do not sit idle waiting for orders
- One agent, one process: Each Anima runs as an isolated OS process and communicates over local IPC (Unix domain sockets, or loopback TCP on Windows), so one failing agent is less likely to take the rest down
- Seven execution engines and multi-model support: Claude Agent SDK, Codex CLI, Cursor Agent CLI, Gemini CLI, Grok Build, models via LiteLLM, and a basic mode. Each Anima can use a different model, and background work such as heartbeat and cron can be routed to a cheaper model
- Web UI, 3D workspace, and voice chat: A dashboard shows the org chart and live status, a Slack-style shared channel called Board hosts discussion, meeting mode gathers several Animas in one room, and browser-based voice chat and a 3D office view are included
- Ten layers of security: Untrusted labeling of external data, five-layer command inspection, a per-agent file sandbox, rate limiting, SSRF mitigation, and fail-closed outbound routing ─ defenses designed on the assumption that autonomous agents hold real permissions
Pricing
| Plan | Price | Highlights |
|---|---|---|
| Open source (the only offering) | $0 (Apache License 2.0) | All features available when self-hosted. No cloud edition or paid tier |
Pricing is current as of August 2026. The software itself is free, but the LLM API usage your Animas consume (Anthropic, OpenAI, Google, and so on) is billed separately. If you are already logged into Claude Code or Codex you can start without an API key, and local models such as Ollama can keep API costs down. Check the official repository for the latest information.
Pros & Cons
✅ Pros
- Apache 2.0 licensed and free, including commercial use, with all data staying in your own environment
- Memory persists across conversations, so you do not have to explain the same context from scratch every time
- The familiar supervisor-subordinate structure means you do not have to design “which agent handles what” in fine detail
- Mixing models lets you tune cost ─ a stronger model for hard judgment calls, a cheap one for routine monitoring
- Security designed from the start around the risks of giving autonomous agents tools
⚠️ Cons
- Self-hosted only. It assumes a Python 3.12+ environment and server operation, with no cloud edition
- Because it is designed to run around the clock, LLM API costs can accumulate if you leave it unattended
- Memory search relies on ChromaDB and similar components, so it needs more resources than a lightweight agent SDK
- It is an individually developed project under rapid iteration, so keeping up with changes takes effort, and some features (such as the Neo4j backend) are experimental
- The documentation is thorough but voluminous, and it takes time to grasp the whole picture
Comparison with Similar Services
| Criteria | AnimaWorks | CrewAI | LangGraph | OpenAI Agents SDK |
|---|---|---|---|---|
| Design philosophy | An organization of autonomous agents | Role-based teams (crews) | Graph-shaped workflows | Lightweight agent SDK |
| Memory | Long-term memory, consolidation, forgetting, automatic recall | Memory features (manual forgetting) | Checkpoints and stores | Session-scoped by default |
| Autonomy | Always on via heartbeat and cron | Human-triggered | Human-triggered | Human-triggered |
| Org structure | Supervisor hierarchy with delegation | Flat role assignment | Depends on your graph design | Handoffs between agents |
| Delivery | Self-hosted (Apache 2.0) | OSS plus paid cloud | OSS plus paid platform | OSS SDK |
Who Is It For
- People who want an “AI staff member” that keeps running daily reports, monitoring, and research, rather than one-off automation
- Anyone tired of rewriting the same background context in every conversation
- Individual developers and small teams who would rather not hand their data to an external SaaS and are fine with self-hosting
- People who want to use Claude, GPT, Gemini, and local models selectively depending on the job
- Developers interested in the design ideas themselves ─ memory, forgetting, and permission isolation in agent frameworks
Summary
AnimaWorks is an attempt to move AI agents from “tools that lose their memory every time you call them” toward “colleagues who remember, and forget.” Org charts, delegation, and overnight memory consolidation start from a different premise than writing workflows one by one, and the difference shows in work that has to keep running continuously. In exchange, it is self-hosted: the API costs of always-on operation and the maintenance burden are yours to carry. A good first step is uv run animaworks demo, which launches a three-person team so you can see the memory and autonomy in action before deciding whether to build your own organization.