WikiDesignCo THE GIGA LIBRARY · ∞ STACKS Request a stack
← The Library
Essay 09 of 12 · Architecture

OpenCog Hyperon: A Clear-Eyed Look at What You Can Actually Use

AtomSpace, MeTTa, Distributed AtomSpace. Research infrastructure that is correct in design and alpha in maturity. How to borrow the ideas without betting the company.

Hero The AtomSpace turning in depth, dense and typed, with links landing on other links. An ALPHA stamp presses onto it, the production-load gauge climbs past the band the single-instance store has proven, and the structure begins to shudder. Admiration and caution, held at once.

There is a specific kind of trap that catches good engineers, and OpenCog Hyperon is one of its purest examples. You go looking for the data model a serious knowledge system needs, the one where a relationship can be the subject of another relationship, where rules are stored the same way as facts, where every assertion carries a degree of belief instead of a bare true or false. You find Hyperon, and it has all of it, designed deliberately and elegantly, by people who have thought about this longer than almost anyone. The model is correct. The temptation is to adopt it. The trap is that the model and the running system are not the same thing, and the running system is alpha.

What you go looking forWhere the AtomSpace already puts it
A relationship that can be the subject of another relationshipA Link is itself an Atom, so a Link takes another Link as one of its arguments
Rules stored the same way as factsOne uniform metagraph, with MeTTa programs held as Atoms in it
A degree of belief on every assertionA truth value on the Atom, carried as a strength and confidence pair

This essay is the counterpart to essay eight, which covered Graphiti, a temporal knowledge graph that is production-mature and structurally limited. Hyperon is the opposite trade: structurally ideal and operationally immature. Reading the two together is the point, because the lesson is not that one is good and one is bad. The lesson is that model maturity and operational maturity are independent axes, and the discipline of a production engineer is to refuse to confuse them. Hyperon gives you the model essay four spent its length reconstructing on a property graph, for free, natively. It also asks you to run research-grade infrastructure underneath a thing your customers depend on, and that is a trade you take with your eyes open or not at all.

Live Two measurements taken on two instruments. Graphiti sits high on operational maturity with a flat model, Hyperon sits far along the model axis on a runtime still in alpha, and the sweep across the field finds no single ranking to read off. The corner both are aimed at stands empty.

So this is the clear-eyed version: what the AtomSpace genuinely is and what runs today, what MeTTa actually does, where the Distributed AtomSpace sits on the line between vision and product, what of the cognitive architecture is real versus aspirational, and which ideas you should borrow even though you cannot yet run Hyperon in production. Every claim is marked for where it sits, runs-now, research-grade, or design-intent, because an assessment that blurs those three is marketing, and the reader this is written for has read enough marketing.

The markWhat it meansWhat this essay files there
Runs nowIn production-adjacent form todayThe classic AtomSpace store
Research-gradeExercised across real projects, without production hardeningPLN, ECAN, MOSES
Design intentArticulated and credible, and not yet builtThe unified cognitive architecture

The AtomSpace is the model done right

Start with what Hyperon gets right, because it is the reason anyone looks. The AtomSpace is a typed, weighted hypergraph, and its single unit is the Atom. There are two kinds. A Node is a symbol with a type and a name. A Link is a typed tuple of other Atoms, and because a Link is itself an Atom, a Link can take another Link as one of its arguments. That one property is the whole thing. An edge can point at an edge, natively, with no reification node, no workaround, no second layer bolted on. The structure that essay four spent four thousand words rebuilding on top of a property graph is the floor of the AtomSpace, present in the type system from the start. It is metagraphs all the way under the hood, and the hood is exactly where the trouble lives.

Live A Link binds two Nodes into a fact. Then a second Link arrives and terminates on that first Link, and the capsule it lands on lights up. Because a Link is itself an Atom it can be an argument, so an edge pointing at an edge sits in the type system from the start.

On top of that structure sits the metadata that makes it a knowledge representation rather than a plain graph. Every Atom carries a type, which gives you schema. Every Atom can carry a truth value, a strength-and-confidence pair rather than a boolean, which is the model essay five built as PLN-lite, present here as the native shape. And in the broader architecture Atoms carry attention values, importance scores that drive what the system thinks about next. Declarative facts, procedural rules, linguistic structure, sensory data, goals, all of it lives in one uniform metagraph of Atoms, which is the property that makes meta-reasoning possible: a system that represents its own rules as Atoms can query and rewrite them like any other knowledge.

Live Three things dock onto one Atom: a type that gives it schema, a truth value carried as a strength and confidence pair, and an attention value that decides what the system thinks about next. On the right, five kinds of content collapse into a single store, which is what lets a system query and rewrite its own rules.

What actually runs, and where the ceiling is

The model is not vaporware, and this is where precision matters most, because there are two distinct code lines and conflating them is the most common error. The classic OpenCog AtomSpace is a mature C++ implementation of a typed hypergraph database that has been exercised for years across robotics, language learning, and biomedical research. It is real, it runs, and it has known numbers: roughly a hundred million Atoms in RAM per instance, a StorageNode that persists on the order of a billion Atoms in about fifty gigabytes, an existing Neo4j backing-store integration for durable persistence, and ProxyNode infrastructure that has federated a small number of networked AtomSpaces. These are runs-now facts, not promises.

Live Three claims, weighted. About a hundred million Atoms live in RAM per instance and about a billion persist in roughly fifty gigabytes, both landing solid with a Neo4j backing store behind them. Federation across machines stops under a ceiling rule, because a small number of networked AtomSpaces is a demonstration and horizontal scale is still being designed.

The ceiling is in that last number. A single instance holding a hundred million Atoms in memory is a real and useful thing, and a billion persisted is respectable. But a handful of federated AtomSpaces is a research demonstration, a long way short of a horizontally-scaled distributed database, and the moment your knowledge system needs to span machines with consistency and availability guarantees, you have left the part of the stack that is proven and entered the part that is being designed. That boundary, between the mature single-instance store and the distributed store that is still taking shape, is the exact line in question, and it is the line the Distributed AtomSpace sits astride.

Scale stepStandingWhat the project reports
Inside one instanceProvenRoughly a hundred million Atoms in RAM, and on the order of a billion persisted in about fifty gigabytes
Across a handful of instancesDemonstratedProxyNode has federated a small number of networked AtomSpaces
Across machines, with consistency and availability guaranteesBeing designedThe Distributed AtomSpace

MeTTa: programs are subgraphs of the same metagraph

The language that runs over the AtomSpace is MeTTa, and it is the second idea worth understanding deeply, because it is genuinely elegant. MeTTa is a pattern-matching and rewriting language: you describe a graph pattern with variables, it matches that pattern against the AtomSpace, and it produces new Atoms or transforms existing subgraphs. That much is a respectable graph query-and-rewrite language. What makes it special is homoiconicity. A MeTTa program is itself represented as a subgraph of the same AtomSpace metagraph it operates on. Code and data are the same substrate. A program is a structure you can store, inspect, query, and rewrite exactly like any fact, which means a program can read and modify another program, and a system can reason about its own behavior because its behavior is written in the same graph it reasons over.

Live A MeTTa rule drawn as what it is, a small subgraph of Atoms inside the same AtomSpace as the data. It matches a pattern, fires, and writes a new Atom back into that same graph. Then a second rule reaches over and reads the first rule's structure, because the first rule is more graph.

The maturity is where the caution returns. What ships in the alpha is a minimal MeTTa interpreter, foundational by its authors' own description, supporting declarative and functional computation over metagraphs and integrated with the distributed store for basic query and rewrite. You can load and run MeTTa scripts, express pattern matching and rewrites, and represent procedures as graph structures the interpreter evaluates. What you do not get is a mature language ecosystem: no settled type system with compile-time guarantees, no real IDE or debugger, a thin standard library, early documentation oriented toward researchers rather than builders, and performance characteristics still under active development. MeTTa today is an alpha-quality interpreter for metagraph rewriting, ideal for research and prototyping, and not a general-purpose production runtime you would put on a critical path. The right move, which the close of this essay returns to, is to borrow the calculus, pattern-match-and-rewrite over a graph, programs-as-data, and implement it in a stack whose runtime you can stand behind.

What the MeTTa alpha shipsWhat is still ahead
Loading and running MeTTa scriptsA settled type system with compile-time guarantees
Pattern matching and rewriting over metagraphsA real IDE and a debugger
Procedures represented as graph structures the interpreter evaluatesA standard library past a thin core
Declarative and functional computationDocumentation aimed at builders
Basic query and rewrite against the distributed storeSettled performance characteristics

The Distributed AtomSpace is the crux

Everything converges on the Distributed AtomSpace, because it is where the gap between the vision and the product is widest, and it is the center of gravity of the whole assessment. The vision is exactly what a serious knowledge platform wants: a distributed, scalable backing store for the metagraph that spans many machines, supports large-scale knowledge management, serves many agents and services concurrently, and carries the episodic memory and the cross-machine reasoning the single-instance AtomSpace cannot. The design documents are clear that this is the goal, and clear that it means going beyond the current limits, the hundred million live Atoms, the billion persisted, the handful of federated instances.

Live The Distributed AtomSpace's stated goal on the left, the bar a mature distributed store already clears on the right, and a measured distance between them labelled ALPHA. That distance is the whole assessment, and it sits on the part of the stack a company would have to depend on.

The reality is that the DAS exists and is implemented in an initial, basic form, enough to run MeTTa programs across a distributed metagraph, and it is used internally for research. The release announces its existence; it does not claim stable performance under load, strong durability guarantees, automatic partitioning, rebalancing, or cluster management. The project's own wiki treats a scalable distributed AtomSpace as an ongoing design topic rather than a solved system. Set against a mature distributed graph database, a Neo4j cluster, JanusGraph over Cassandra, a managed cloud graph service, the DAS is an early research vehicle, with evolving APIs, changing on-disk formats, and limited operational tooling. This is the precise meaning of gorgeous research infrastructure that is not yet production. The DAS is where you would have to bet the company, and it is the one part of the stack least ready to be bet on.

The cognitive layer is mostly somewhere else

The reason Hyperon exists at all is the layer above the metagraph: a full cognitive architecture, probabilistic logic in PLN, attention allocation in ECAN, program evolution in MOSES, all sharing one AtomSpace and coordinated through MeTTa into something aimed at general intelligence. This is the ambition that makes the project compelling, and it is also the part where the gap between what is written about and what is running is largest, so it needs the plainest accounting of all.

Live PLN, ECAN and MOSES, each a working module with research mileage, each addressed to classic OpenCog. The arrows toward the unified Hyperon home stop at the boundary, and the slots inside that home stay drawn as outlines.

The components are real as a body of research code, and that distinction is the whole point. PLN, the probabilistic logic system that reasons over truth values, ECAN, the attention mechanism that allocates importance across Atoms, MOSES, the evolutionary program-learning system, all exist and have been exercised in classic OpenCog across real projects. But the Hyperon documentation and the alpha release focus on the foundational infrastructure, the AtomSpace metagraph, the DAS, and MeTTa, and are explicit that the higher cognitive layer, while it is the entire purpose, has not been the focus of the Hyperon-oriented work so far. Some early integration surely exists in labs and internal repositories. What does not exist is a complete, production-quality cognitive stack you can pick up and run. The story of cognitive parts working in concert, PLN and ECAN and MOSES and neural modules all sharing one metagraph through MeTTa, remains a research program and a design goal, still some distance from an off-the-shelf architecture, and an assessment that presents it as the latter is selling something.

Where the Hyperon-oriented work has goneWhat sits there
The foundational infrastructureThe AtomSpace metagraph, the Distributed AtomSpace, and MeTTa
The layer that is the project’s entire purposePLN, ECAN and MOSES in concert through one shared metagraph, which the alpha is explicit has not been the focus so far
What that leaves a builderResearch code and reference designs, with a complete cognitive stack you can pick up and run still ahead

The ledger, stated plainly

Put the whole stack on one axis and the picture resolves. Three columns: runs in production-adjacent form today, research-grade and exercised but not productized, and design-intent not yet built. The classic AtomSpace store sits in the first column, a real typed-hypergraph database with known capacity and a Neo4j backing option. PLN, ECAN, and MOSES sit in the second, real algorithms with research mileage but no production hardening, and living in classic OpenCog rather than ported into Hyperon. The DAS and the minimal MeTTa interpreter sit between the second and third columns, implemented in initial form and usable for prototypes, with load-readiness still ahead of them. And the unified cognitive architecture, the distributed metagraph at scale, the self-improving MeTTa programs, sit in the third, articulated and credible and not yet here.

Live Every component sorted once. The classic AtomSpace store lands in RUNS NOW, PLN and ECAN and MOSES in RESEARCH-GRADE, the unified architecture in DESIGN INTENT, and the Distributed AtomSpace and the minimal MeTTa interpreter come to rest on the boundary between the last two.

The ledger is the answer to the question the essay's title asks. What you can actually use today is the classic AtomSpace as a reference store and the published algorithms as reference designs. What you can experiment with is the Hyperon alpha, the metagraph plus DAS plus MeTTa, as a research playground. What you cannot do is run any of it underneath a customer workload and sleep at night. None of that is a criticism of the project, which is doing exactly what an ambitious research program should. It is a statement of where the line is for someone who owes their employer uptime.

The verdictWhat it covers
Use todayThe classic AtomSpace as a reference store, and the published algorithms as reference designs
Experiment withThe Hyperon alpha, the metagraph plus DAS plus MeTTa, as a research playground
Keep off the critical pathAny of it running underneath a customer workload

Borrow the ideas, not the dependency

The productive stance is to treat Hyperon as a reference design and steal its best ideas onto infrastructure you can operate. Five ideas are worth taking, and each has a mature-substrate implementation. The first is the metagraph model itself, edges that point at edges, which you build on a property graph through the reification pattern essay four laid out: promote a relationship to a node and let other edges point at it. You lose the native elegance and you keep the running database. The second is programs as graphs: store rules, workflows, and inference templates as graph structures in your own store, so your system's logic is itself queryable and explainable, while you keep your real runtime in a conventional language.

Live Five ideas lift off the dimmed Hyperon slab and land on infrastructure you already run. Reification carries the metagraph, your own store carries rules as data, a small DSL carries pattern-match and rewrite, two numeric fields carry truth values, and importance becomes a cache and scheduling heuristic.

The third idea is pattern-matching and rewriting as the core computation over your knowledge, MeTTa's calculus, implemented as a small embedded DSL or a Datalog-style engine over your graph rather than as MeTTa itself. The fourth is truth values as first-class metadata, a strength-and-confidence pair on every edge instead of a boolean, which essay five built as a working PLN-lite and which you can carry as two numeric fields plus a handful of combination rules, no full probabilistic logic engine required. The fifth is attention as resource management, ECAN's importance scores reframed plainly as what they operationally are, a cache and scheduling heuristic, frequently-used facts kept hot, hot rules tried first, eviction and pre-computation driven by importance. None of the five requires running Hyperon. All five are the good ideas Hyperon organized, and all five run on infrastructure that stays up.

The ideaWhat Hyperon gives itHow you run it on a substrate you operate
The metagraph modelLinks that take Links as arguments, native to the type systemReification on a property graph: promote a relationship to a node and let other edges point at it, the pattern essay four laid out
Programs as graphsMeTTa programs stored as Atoms in the same metagraph as the dataRules, workflows and inference templates stored as graph structures in your own store, so the logic is queryable and explainable
Pattern-match and rewriteThe core computation over the whole knowledge baseA small embedded DSL or a Datalog-style engine over your graph
Truth valuesA strength and confidence pair on every AtomTwo numeric fields plus a handful of combination rules, which is the PLN-lite essay five built
AttentionECAN importance scores allocated across the AtomSpaceCache and scheduling: frequently-used facts kept hot, hot rules tried first, eviction and pre-computation driven by importance

When to look again

Live Two gates and one rail, at the same time. Each gate rests at alpha while a probe runs toward a stable, non-toy release and falls back short of it. The borrowed model keeps moving into a store you can operate, which is what makes the waiting affordable.

The clear-eyed close is a posture the reader can hold while the picture keeps moving. Hyperon is a thing to track, and tracking it means watching the maturity gap close or fail to close rather than settling the question today. The discipline that keeps you out of trouble is the one this essay opened with: model maturity and operational maturity are independent, and you can admire the first while refusing to depend on the second. Take the model now, through reification and first-class truth values and a rewrite DSL, on a store you can operate, which is the bridge essay four built and essay eight's Graphiti is one production answer to. Watch the runtime, and revisit Hyperon when its Distributed AtomSpace and its MeTTa stack reach a stable, non-toy release with evidence from deployments that are not demonstrations. That day may come, and the model will be waiting, already correct. Until it does, the company runs on infrastructure that is boring and up, and borrows its ideas from research that is beautiful and early.

The next essay turns from the substrate to retrieval: how an agent actually pulls the right knowledge out of a graph this rich, where hybrid search meets the metagraph, and why the structure that took all these essays to build finally pays off at query time.


END OF ESSAY 09 · CONTINUE TO ESSAY 10 →