Hypergraph vs Hyperbolic Graph for AI Memory

An agent's memory has shape, and the shape decides the structure.

A long-lived agent does not store a flat list of strings. Its knowledge bends in two ways at once. It is hierarchical — "spaniel" sits under "dog" under "animal"; a session rolls up into a topic, a topic into a domain. And some of its facts are genuinely n-ary — "in session X, agent A used tool T on resource R and got outcome O" binds five things in one relation, not a pair.

Those two pressures point at two different ways to generalize a graph, and they get confused constantly because both start with "hyper-": a hyperbolic graph curves the space; a hypergraph grows the edge. They are not variants of each other. Pick the wrong one and you optimize for a problem your memory does not have.

TL;DR

• Hyperbolic graph — pairwise edges in negatively curved space, for hierarchy. This is the axis Mnemoverse bets on.
• Hypergraph — one edge joins many vertices, for n-ary relations. A real, open question we have not taken; our graph is pairwise today.
• The two axes are orthogonal, so a hyperbolic hypergraph (both at once) also exists.
• Our tensor-hyperbolic graph (THG) is a hyperbolic graph, not a hypergraph — and it is research-stage, not in production.

A hyperbolic graph keeps ordinary two-vertex edges but places the graph in curved space. A hypergraph keeps flat space but lets one edge join three, five, or ten vertices at once.

Two axes, not one

An ordinary graph is two commitments at once: every edge joins exactly two vertices, in ordinary (Euclidean) space. Loosen either commitment and you get a different structure — and a different answer to "what does this buy an agent's memory?"

Structure	An edge joins	Lives in	What it buys an agent's memory	Anchor
Plain graph	2 vertices	Euclidean	general pairwise links between memories	—
Hypergraph	many vertices	Euclidean	n-ary facts as one relation (who-did-what-with-what)	HyperGraphRAG (arXiv:2503.21322)
Hyperbolic graph	2 vertices	hyperbolic	hierarchy with low distortion (concept trees, roll-ups)	HGCN (arXiv:1910.12933)
Hyperbolic hypergraph	many vertices	hyperbolic	n-ary and hierarchical at once	H2GNN (arXiv:2412.12158)

The rest of this note walks the two middle rows — and lands on which one Mnemoverse picked.

The hierarchy pressure → hyperbolic graph

Start with the uncontroversial part: agent memory is mostly hierarchical. Concepts nest, sessions roll up into topics and domains. When we let the memory engine grow on real conversation data, it self-organizes into a power-law structure — a few hub concepts accumulate most of the connections, the same shape you see in citation graphs and Wikipedia (Building Memory That Scales). Hierarchy is not a design we imposed; it is what knowledge does.

Euclidean space embeds deep hierarchies badly — a tree's branching count explodes exponentially while flat volume grows only polynomially, so nodes pile up and distances lie. Hyperbolic space fixes this by construction: volume grows exponentially with radius, matching how a tree branches. This is established prior art and we claim none of it — a weighted tree embeds in the hyperbolic plane with arbitrarily small distortion (Sarkar, 2011), reaching near-perfect precision in as few as two dimensions (Sala et al., 2018), and learned Poincaré embeddings beat Euclidean ones on hierarchical data (Nickel & Kiela, 2017).

A hyperbolic graph puts this to work while leaving the edges alone: pairwise edges, laid out in the Poincaré ball instead of flat space. So the cue is simple — if the dominant relation in your memory is "X is part of / a kind of Y," you want the hyperbolic axis, and you change the space, not the edge. Hyperbolic graph convolutions cut link-prediction error by up to 63.1% in ROC AUC over state-of-the-art (Euclidean) GCN baselines (Chami et al., 2019, HGCN) — evidence the geometry helps in general. That figure is theirs, on their benchmarks; it is not a measurement of Mnemoverse's memory engine and should never be read as one.

The practical wall: nobody ships native hyperbolic embeddings

The math advantage is real — and it is the wall Mnemoverse actually hit. Every mainstream production text-embedding model — OpenAI text-embedding-3, Cohere Embed, Voyage, Gemini Embedding, the open sentence-transformer family — emits Euclidean vectors compared by cosine similarity; we could not find a widely available model that natively outputs hyperbolic (Poincaré-ball) vectors. Native hyperbolic NLP is still research-stage — Poincaré GloVe, hyperbolic fine-tuning (HypLoRA), and the first billion-parameter fully-hyperbolic LLMs (HELM, NeurIPS 2025). And the geometry cannot be bolted on after the fact: an encoder trained in Euclidean space never learned hyperbolic structure, and the usual fix — read a Euclidean vector as a tangent vector at the origin and apply the exponential map — chains maps that cancel back to an approximately Euclidean transformation (the HypLoRA analysis). That is why the shipping engine projects Euclidean embeddings onto the manifold and calls the result approximate: a workaround, not native hyperbolic training.

The n-ary pressure → hypergraph

Now the other axis. A hypergraph leaves the space flat and changes what an edge is: a hyperedge is a set, joining many vertices in a single relation. As the HyperGraphRAG authors put it, "each edge in an ordinary graph connects only two entities, limiting their ability to represent the n-ary relations" (Luo et al., NeurIPS 2025).

For agent memory this is the who-did-what-with-what-and-got-what problem. "Agent A used tool T on resource R in session X and got outcome O" is one fact about five entities. A pairwise graph has two ways to stand in for it, and they are not equal: explode the relation into a clique of edges and you lose which links belong together — a five-clique reads the same as five unrelated pairs — or add one node that stands in for the relation and link all five participants to it, which keeps the membership intact at the cost of a little indirection. A true hyperedge holds the relation as a single first-class object. So the second cue: if your facts routinely bind three or more entities into one irreducible relation, you are feeling n-ary pressure, and you change the edge, not the space.

This is a real, open question for memory, and we have not taken it. Mnemoverse's graph is pairwise today. HyperGraphRAG shows the n-ary job is real and valuable — a different job from hierarchy, not a worse one.

A cheap, lossless bridge to n-ary

You do not actually need a hypergraph to store an n-ary fact on a pairwise graph. Add one node that stands in for the relation and link every participant to it — the "star expansion" (or incidence / Levi) graph in graph theory, the same construction as a W3C n-ary-relation node, RDF reification, and the "event node" of formal semantics. It is lossless for membership, not a lossy approximation: the encoding is a bijection, so which entities belong to which relation is always recoverable — unlike clique expansion, which wires participants pairwise and genuinely cannot tell a five-way relation from five separate pairs. The price is concrete and small: one extra node per relation, a hop of indirection (participants sit at distance two, not directly adjacent), and a plain graph model will not exploit hyperedge-level structure the way a purpose-built hypergraph network would. So the n-ary axis is not a capability wall for a pairwise engine like ours — it is a deliberate simplicity choice with a known, cheap bridge if the workload ever demands it.

Both at once → hyperbolic hypergraph

Because the axes are orthogonal, you can loosen both commitments. A hyperbolic hypergraph has many-vertex hyperedges and lives in curved space — the natural home for knowledge that is n-ary and hierarchical, like a five-way event that also sits at a particular depth in a taxonomy. It is an active research line: hyperbolic hypergraph networks operate in hyperbolic space "to capture the tree-like hierarchy" while keeping the hyperedge's n-ary structure (Li et al., 2024, H2GNN). Note this is distinct from a plain hyperbolic graph — it changes the edge too.

Where THG sits — and the caveats

Our tensor-hyperbolic graph (THG) is a hyperbolic graph — the third row, never the fourth. Every edge connects exactly two nodes; there are no hyperedges anywhere. What is unusual is the attribute: instead of a scalar weight, each edge carries a small rotation tensor — an O(d) rotation read off the two endpoints' positions — with the whole graph embedded in the Poincaré ball. THG is the geometric substrate for our Semantic Level of Detail (SLoD) operator, and it has no learnable parameters (μ = 0): nothing is fit; the geometry is read off the embedding. The "tensor" is the edge's attribute, not a license to connect many vertices.

Two caveats. First, "tensor-hyperbolic graph" is our own framing for this synthesis, not an established term — and the THG acronym appears elsewhere in the literature for unrelated work, so treat the name as ours. Second, none of this is in the production engine: the shipping engine projects Euclidean embeddings onto the manifold, which is approximate; native hyperbolic embedding is the missing piece, and THG and SLoD are research-stage (Building Memory That Scales). What is shipping points the same way from a different angle — the production graph self-organizes those power-law hubs as it grows. That the structure is hierarchical is observed; that hyperbolic geometry is its right native home is the bet we have not yet cashed.

So THG is a bet on the hyperbolic axis, not a claim to have solved memory's geometry, and the hypergraph axis stays open. The field guide fits on one line: hierarchy → hyperbolic graph; n-ary → hypergraph; both → hyperbolic hypergraph. The only outright mistake is to treat the two words as synonyms — they answer different pressures, and an agent's memory feels both.

Common questions

What is the difference between a hypergraph and a hyperbolic graph? They generalize a graph along two orthogonal axes. A hypergraph changes the edge — one hyperedge joins many vertices at once, for n-ary relations. A hyperbolic graph keeps ordinary pairwise edges but places the graph in hyperbolic (negatively curved) space, which fits hierarchies with low distortion. The two axes can be combined.

Which one does an AI agent's memory need? Both pressures show up. Memory is hierarchical (a session rolls up into a topic, a topic into a domain), which is the case for hyperbolic space; and some memories are genuinely n-ary ("agent A used tool T on resource R and got outcome O"), which is the case for hypergraphs. They answer different questions. Mnemoverse bets on the hyperbolic axis (THG); the hypergraph axis is an open question we have not taken.

Is a tensor-hyperbolic graph (THG) a hypergraph? No. THG is a hyperbolic graph — every edge connects exactly two nodes (it carries a rotation tensor instead of a scalar weight), embedded in the Poincaré ball. It is not a hypergraph, and "tensor-hyperbolic graph" is our own framing rather than a field-standard term.

Sources

• Luo et al., HyperGraphRAG: Retrieval-Augmented Generation via Hypergraph-Structured Knowledge Representation, NeurIPS 2025 — arXiv:2503.21322
• Li, Shi, Qiao, Zhang & Jin, Hyperbolic Hypergraph Neural Networks for Multi-Relational Knowledge Hypergraph Representation (H2GNN), 2024 — arXiv:2412.12158
• Chami, Ying, Ré & Leskovec, Hyperbolic Graph Convolutional Neural Networks (HGCN), NeurIPS 2019 — arXiv:1910.12933
• Nickel & Kiela, Poincaré Embeddings for Learning Hierarchical Representations, NeurIPS 2017 — arXiv:1705.08039
• Sala, De Sa, Gu & Ré, Representation Tradeoffs for Hyperbolic Embeddings, ICML 2018 — arXiv:1804.03329
• Sarkar, Low Distortion Delaunay Embedding of Trees in Hyperbolic Plane, Graph Drawing (GD) 2011 — doi:10.1007/978-3-642-25878-7_34
• Noy & Rector (eds.), Defining N-ary Relations on the Semantic Web, W3C Working Group Note, 2006 — w3.org/TR/swbp-n-aryRelations
• Hyperbolic Fine-Tuning for Large Language Models (HypLoRA), 2024 — arXiv:2410.04010
• HELM: Hyperbolic Large Language Models via Mixture-of-Curvature Experts, NeurIPS 2025 — arXiv:2505.24722

Related

• Tensor-Hyperbolic Graphs — our worked example of a hyperbolic graph for memory
• Semantic Level of Detail — the operator built on that geometry
• Building Memory That Scales — why production does not yet run native hyperbolic geometry

Edward Izgorodin — last reviewed June 2026.