Raven: A Memory-First, Self-Improving Agent Harness Built on EverOS

What Raven Actually Solves Today

Raven is positioned as a self-improving agent harness. The README calls it the harness of harnesses: as a Host Agent it brings built-in agents (Raven-Research, Raven-Code, Raven-Design and Raven-Oncall) and third-party agents together, generating DAGs to orchestrate them on complex tasks. Its package describes it as an AI-native command line agent with memory, proactivity, context control, and skill evolution. The target user is an engineer or researcher who wants to run agents for hours or days, inspect what the agent did, and let it improve across sessions. The memory component is not a bolt-on; it is EverOS, which carries memory and context across sessions. This is not a library you import into a script. It is a harness you run from a terminal, with a terminal UI (ui-tui/) and a web UI (ui-web/). The distinction matters: Raven aims to be the environment where agents live, not a function you call.

The Mechanism: TUI, Onboarding, and the Six-Config Wizard

The engine is a Python package that requires Python 3.12 or newer, with dependencies including typer, litellm, rich, prompt-toolkit and mcp. The repository also carries ui-tui/ and ui-web/ for its terminal and web interfaces. The Dockerfile notes that the engine binds its HTTP surface to 127.0.0.1, which is why the image runs nginx beside the engine in one container.

Getting It Running: Commands and Config Keys

The repository root ships install.sh and install.ps1. For containers, the Dockerfile documents `docker build -t raven:local .` and `docker build -t raven:local --build-arg RAVEN_EXTRAS="channels,tools,sandbox" .`, and points to docker/docker-compose.yml for running it. The Makefile also defines docker-build, docker-up and docker-down targets.

Deep Research: MiroThinker and the Cost Gate

The repository description credits MiroThinker with powering Raven's deep research and reasoning, and the built-in Raven-Research agent covers research work. The limitation is that this path depends on MiroThinker, a separate project, not something Raven controls.

Tracing: The Local Audit Trail

The Limits: Pre-Alpha, the Harness of Harnesses Gap, and Benchmark Caveats

pyproject.toml classifies the package as "Development Status :: 3 - Alpha". The README's showcase results are Raven's own reports: for example, the RSI run completed 172 nanochat training runs across 7 rounds and reduced val_bpb by 5.8% within the same 20-minute, single-GPU budget. The wrong tool case is clear: if you need a stable, production-grade agent framework with a mature API, Raven is not it. Recent releases have come every one to three weeks (v0.2.4 on 2026-10-03, v0.2.5 on 2026-10-20, v0.2.6 on 2026-10-27), which suggests rapid iteration but also instability.

Alternatives and the Real Difference in Approach

For a practical comparison, the closest alternative is a framework like LangGraph or AutoGen, which also build agent loops with memory and tool use. The difference in approach is structural. LangGraph gives you a graph-based state machine where you define nodes and edges explicitly; you control the flow. Raven is a harness that wraps the flow for you, generating the DAG that orchestrates its agents. You do not write a graph; you configure a system. That is a trade-off. If you need fine-grained control over the agent's decision points, a graph framework is better. If you want a ready-made loop with tracing and memory built in, Raven is closer to that. This comparison is inference, but the design choices are visible: Raven ships built-in agents, terminal and web UIs, and MiroThinker-powered deep research, none of which are typical in a graph library.

Maintenance, Upgrades, and License

The project is Apache-2.0 licensed, which is permissive for commercial use, and the repository carries CONTRIBUTING.md, CODE_OF_CONDUCT.md and RELEASING.md. The release cadence is active: v0.2.6 was released on 2026-10-27, with v0.2.5 a week earlier and v0.2.4 on 2026-10-03. That suggests a team that is shipping quickly, but it also means you should expect breaking changes. The cost of maintenance is that you will need to track releases and test your workflows against each new version. For an alpha project, that is the expected trade-off. The license is permissive, but the project depends on MiroThinker for deep research, which is a separate service with its own terms; the README does not detail those terms, so you should check them before relying on that path.

Editorial conclusion

Adopt Raven if you want a memory-first agent harness with long-term memory through EverOS, built-in agents and MiroThinker-powered deep research, and you can tolerate alpha instability. Do not adopt it if you need production reliability, a stable API. Before using it, verify that your LLM provider is supported by the onboarding wizard, that your sandbox or execution location meets the security requirements for running untrusted code. Pin your version if you depend on a specific behavior, since releases arrive every few weeks.

Frequently asked questions

Official sources

1. Official documentation
2. Official README
3. Project repository
4. Release notes