Hyperon is the artificial intelligence (AI) framework developed by Ben Goertzel and his team. The current version of the Hyperon white paper is titled “Hyperon: The Open-Source Infrastructure for Artificial General Intelligence” (2026). See also this shorter business-oriented summary.
This is my first followup to “ELI5: Hyperon for complete idiots (part 1),” written in December of last year. Since then I’ve been thinking about which part of Hyperon I should cover next. I guess I should first say something about the Hyperon programming language MeTTa.
MeTTa in a (very small) nutshell
The MeTTa language (the name was chosen as short for Meta Type Talk, then retained for its associations with Buddhist philosophy), has been created by Ben Goertzel and collaborators working on the Hyperon project under SingularityNET and the ASI Alliance. MeTTa is designed as the foundational “language of thought” for artificial general intelligence (AGI) systems based on Hyperon.
MeTTa is built to represent and manipulate knowledge inside Hyperon’s metagraph (the Atomspace, see part 1). MeTTa programs work by querying and rewriting portions of the Atomspace. The language deliberately blends functional and logical styles - its own style can be roughly described as a mix of Lisp and Prolog styles. Knowledge and executable procedures can live side by side.
The basic units of MeTTa are atoms - fundamental units that can represent objects, concepts, processes, functions, or relationships. Expressions nest atoms inside parentheses, variables begin with a dollar sign and enable pattern matching, and grounded atoms link to external code or data (e.g. the number 1 is a grounded atom).
Program statements act like rewrite rules or function definitions. Pattern matching and search form the core of evaluation. Since code and data share the same representation, programs could eventually inspect and rewrite themselves at run time, supporting the kind of self-modification that advanced artificial general intelligence (AGI) systems are expected to need. MeTTa is nondeterministic - several different evaluations can be valid at the same time, and the language returns all of them.
There’s no “MeTTa for complete idiots” explainer that I know of (note to self: maybe I should write one for the next episode of this series). If you read MeTTa code (see below for examples), you’ll be treated with a deluge of nested parentheses that will make you think of Lisp. In fact, MeTTa resembles Lisp through its parenthesized prefix expressions
(e.g. (* 2 (+ 3 4)) evaluates to 14).
This notation may feel unusual to amateur programmers, but it is based on solid computer science concepts and after a while begins to feel natural. The site metta-lang.dev has a playground where you can practice evaluation of MeTTa expressions.
Other similarities between MeTTa and Lisp are support for higher-order functions and recursion, and homoiconicity - this means that code and data share the same representation, and programs can manipulate other programs as data. There are also similarities with Prolog. MeTTa offers rich functional constructs, full access to its own code, and direct integration with external systems such as neural networks. These features position MeTTa as a meta-language intended both for human developers writing AGI components and for the learning algorithms that may eventually rewrite those components. MeTTa is also usable as a programming language for smart contracts in blockchain systems.
Future episodes of this series will include more about MeTTa and other parts of Hyperon. But now I want to say something about a significant new development in Hyperon: OmegaClaw.
Here come the Lobsters!
“OmegaClaw is the first thing out of our long-brewing Hyperon AGI research initiative that the average person can actually sit down and chat with in a simple way,” says Ben Goertzel. “It’s basically an attempt to take an OpenClaw-style cognitive architecture, re-implement the bones of it in MeTTa - which, recall, was designed for self-modification - and then bolt it onto the Hyperon AtomSpace as long-term and working memory, so the agent has a real symbolic substrate it can reflect on, revise, and reason over.”
All started with the launch of OpenClaw in late 2025 and the rapid rise of OpenClaw and Moltbook (a social network for bots) in early 2026. Peter Diamandis called it the Lobster revolution. “The lobsters (yes, that’s the mascot, inspired by Charles Stross’s Accelerando),” he said, “have arrived, and they’re not leaving.” In Stross’ novel, we first meet them when researchers are uploading Californian spiny lobsters into cyberspace, “one neuron at a time,” and then again and again as the Singularity unfolds on this planet and in the vast universe out there..
“Super fun!” posted Ramez Naam about OpenClaw, “Like a Charlie Stross novel!” Naam added “Not a sign of AGI.” But using this technology in the pursuit of AGI in the Hyperon framework is exactly what Goertzel and his team want to do.
OmegaClaw was all over the recent AGI-26 conference. See especially Goertzel’s talk on “The path to AGI/ASI: What we need to get there” at the beginning of the last day. See also, in the latest part of the same last day, Khellar Crawford’s remarks on OmegaClaw. “OmegaClaw came about when Ben suggested to the team ‘can we implement OpenClaw in MeTTa?’,” he said. Development has been fast since then.
Crawford’s remarks were followed by a simplified but delicious hide-and-seek demo with a robot called Codey, powered by OmegaClaw. The demo was a sign of things to come: “When OmegaClaw integrates with Codey,” notes GeekWire, “the robot should build ongoing relationships and remember every conversation, rather than starting fresh every time.”
What is OmegaClaw?
OmegaClaw is “a mix of Hyperon and LLMs” (large language models), said Goertzel at AGI-26.
A simplified picture of an OmegaClaw agent is a continuous MeTTa loop that teams up with an LLM. More precisely, the LLM is one component, but intelligence lives in the whole system: a persistent knowledge base implemented in MeTTa, a multi-tier memory that survives across sessions, and a context-building process that actively assembles knowledge before forwarding it to the LLM.
What sets the agent apart from a plain chatbot is this integrated memory and knowledge representation. It accumulates experience over time, revises beliefs when new evidence comes in, and can extend its own abilities by adding tools and rules as it goes. The memory and knowledge representation are what make this reasoning possible across sessions. The agent develops a sense of continuity and gradually builds a character. Users talk to it through messaging channels like Telegram.
An instance of OmegaClaw itself, and human participants in a group chat, helped me refine this short description.
Testing, testing 1 2 3
I wanted to experiment so I installed OmegaClaw-Core from GitHub. On my MacBook Pro (Apple M1 Max, latest operating system Tahoe 26.5.2), the installation procedure worked as explained in the main repository page, without problems. I’m running OmegaClaw in Docker. The installation procedure is simple for users with a certain familiarity with computers, but I wouldn’t say it is at everyone’s fingertips yet. Bear in mind, though, that this is still experimental and improvements are likely to come soon on all fronts.
After installing OmegaClaw, one has to choose which LLM to use. This is an important choice, as it impacts quality and cost (as you may have guessed, best is more expensive). I started with Claude models, but soon realized that continuing with Claude models would have been too expensive, so I switched to GLM 5.2 via OpenRouter.
Another choice to make is how to communicate with the agent. Of the options listed by the system after installation, I choose a Telegram bot (for no other reason that I use Telegram already). It is simple: one has to create and configure a bot using the master bot BotFather, then pass the bot token to OmegaClaw. I called my bot Wunchy (full name WunchyBot) because I wanted a name out of Accelerando (the Wunch is a collective of nasty cybernetic aliens originating from the original digitized lobster minds).
I started chatting with Wunchy. First, I asked Wunchy to explain OmegaClaw to me, locate the important source files and answer questions. Wunchy told me how to edit the loop.metta file (take a look if you want to see an example of working MeTTa code) to reduce costs by reducing the frequency of LLM prompts.
Then I asked Wunchy to help me solve an issue with the Telegram bot: Wunchy only talks to me and doesn’t talk to others (doesn’t even see their messages), even after I added Wunchy to a Telegram group chat with other users. This took some time and wasn’t successful - after many trials I put everything back as it was. One of the developers told me that the implementation of the Telegram interface in this version of OmegaClaw is minimal and doesn’t support multi-user chat. I look forward to testing new versions of OmegaClaw.
One issue that I found is that, even after switching from Claude to GLM, running Wunchy is still expensive for a hobby project. I think this is due to the history file (the file history.metta in the memory directory is an empty placeholder, but you can find the active one in the Docker workspace). A tail of the history file (configurable in memory.metta) is passed to the LLM at each prompt (or more frequently depending on the configuration), and token usage goes up. I’m sure future releases of OmegaClaw will be optimised to save costs. There are models cheaper than GLM, but the quality of the underlying LLM drives the overall performance, so the LLM must be at least good enough.
An OmegaClaw agent, dedicated to finding and summarizing important news about AI and AI-adjacent sci/tech, will soon be integrated into Mindplex.