MnemoverseDocs

Appearance

Edward Izgorodin·Published Apr 6, 2026Updated Jun 22, 2026

Tensor-Hyperbolic Graphs for AI Memory ​

A hierarchy of knowledge branching outward through hyperbolic space — toward the boundary there is exponentially more room.

You store knowledge that has shape. "Spaniel" sits under "dog" sits under "animal"; a function lives in a module lives in a package. Most vector-based memory systems flatten that shape away, storing everything as a list of points where the only relationship an edge can express is distance.

Here is the flag we are planting, stated plainly so it can be argued with: we bet that AI memory is intrinsically hyperbolic — that knowledge has the shape of a tree-like field, and that a single graph laid out in hyperbolic space can, in the limit, naturally cover the whole of it. That sentence is a hypothesis, not a result. The rest of this note earns the part we can defend and is ruthless about the part we cannot.

A tensor-hyperbolic graph (THG) is a knowledge graph embedded in the Poincaré ball in which each edge carries a tensor — an O(d) rotation computed from its two endpoints' positions — rather than a single scalar weight. It is the geometric substrate for our Semantic Level of Detail (SLoD) operator. "Tensor-hyperbolic graph" is our framing for this synthesis; the acronym appears elsewhere in the literature for unrelated work.

TL;DR

  • The bet: AI memory is intrinsically hyperbolic, and a tensor-hyperbolic graph is the structure that covers the knowledge field. This is a labelled conjecture, not proven.
  • The proven core (narrow): the connection Laplacian is positive semi-definite, and a tree is the unique zero-tension topology — so any tension at all certifies a departure from a single clean hierarchy.
  • The honest gaps: loss-minimization self-organization was tried four ways and disproved; three formulas were retracted along the way; the grand thesis is unmeasured.
  • Not in production. The shipping engine projects Euclidean embeddings onto the manifold — approximate. Native hyperbolic embedding is the missing piece.

Why we bet memory is hyperbolic ​

Start with what is not controversial. Knowledge is hierarchical: concepts contain subconcepts, topics contain subtopics. Euclidean space embeds deep hierarchies badly, because its volume grows polynomially while a tree's branching grows exponentially. The Poincaré ball — a model of hyperbolic space — grows volume exponentially, matching that branching almost exactly.

This is established prior art, and we claim none of it. A weighted tree embeds in the hyperbolic plane with distortion that can be made arbitrarily small (Sarkar, 2011); trees reach near-perfect precision in as few as two dimensions (Sala et al., 2018, arXiv:1804.03329); learned Poincaré embeddings beat Euclidean ones on hierarchical data (Nickel & Kiela, 2017, arXiv:1705.08039), with later work extending this to graph networks and mixed-curvature spaces (Ganea et al., 2018; Chami et al., 2019; Gu et al., 2019). On link prediction, hyperbolic graph convolutions cut error by up to 63.1% in ROC AUC over Euclidean baselines (Chami et al., 2019) — evidence the geometry helps in general, not a measurement of our system.

Our bet takes one step past all of this. Hierarchy embeds in hyperbolic space; we are betting that knowledge memory as a whole does — that a single hyperbolic graph is a navigable substrate the whole field can be read from, not just clean taxonomies. That last step is the conjecture. We have not proven it, and we will not pretend the prior art does.

What the tensor adds ​

In a standard hyperbolic graph an edge is a number. In a THG, an edge carries a tensor: an O(d) rotation — a rigid turn in d dimensions — that aligns one node's local frame with its neighbour's. Think of it as a small instruction for how to turn as you move from one concept to the next, not just how far. Both the rotation and the edge's strength (a conformal weight) are computed directly from where the two nodes already sit — no learnable parameters (μ = 0). Nothing is fit; the geometry is read off the embedding. The library exposes this through a single GraphStore class that is signals-only: it carries the canonical edge map and emits measurements, it does not train weights.

Why bother?

Because a rotation composes, and a number does not.

Walk a loop in the graph, multiply the rotations you pass through, and you return slightly turned. That leftover turn is a signal a scalar weight can never carry — and it is the thing this whole structure is built to read.

We call that residual Tax (tension). Here is the part we can defend: in our internal math notes we prove that the underlying connection Laplacian is positive semi-definite, and that a tree is the unique topology with zero Tax. Every cross-link that breaks a clean hierarchy — the edge that makes "bat" both a mammal and, loosely, a flier — shows up as nonzero tension. Within our internal formulation, then, Tax is a parameter-free measure of how far a graph departs from a single hierarchical embedding. (That it is an exact spectral residual is a separate, still-open conjecture.) For a triangle the holonomy — the leftover turn from walking around it — even has a closed form, 2√2·|sin(Area/2)|, a correction we documented after an earlier formula was wrong.

What is a bet, and what is proven ​

A fringe idea is only credible if it is honest about its own status. So, the ledger:

  • Proven (stated in our internal math notes): the connection Laplacian is PSD; a tree is the unique zero-tension topology; the triangle holonomy is 2√2·|sin(Area/2)|. We frame these as stated-and-proven internally, not as independently established results.
  • Conjectured / open (labelled as bets): that AI memory is intrinsically hyperbolic; that one graph covers the whole knowledge field; that Tax is an exact spectral residual; that detail boundaries correspond to local tension gradients; that Tax increases monotonically with dimension; that matryoshka multi-resolution holds (the projections provably do not commute — they only approximately agree). These are geometrically attractive and not proven.
  • Retracted along the way: three formula corrections, and an entire approach — building the structure by loss minimization — which we tried four ways and disproved. The working method is deterministic placement plus Tax as a measurement, not optimization.
  • Not in production: the shipping engine projects Euclidean embeddings onto the manifold, which is approximate. Native hyperbolic embedding is the missing piece.

What we have measured ​

The numbers worth stating come from our published SLoD paper, on synthetic and benchmark data — not from the THG library in production:

  • On synthetic hierarchies (HSBM, 1024 nodes), macro abstraction boundaries are recovered at ARI = 1.00; meso boundaries at ARI ≈ 0.89 [0.86, 0.92], at r = 200 (Izgorodin, 2026, arXiv:2603.08965).
  • On the WordNet noun hierarchy (82K synsets), detected boundaries align with taxonomic depth at Kendall τ = 0.79 (same paper).

Inside the THG library, an internal micro-benchmark on a 15-node test tree preserves hierarchy perfectly and tracks pairwise distance closely — a single run, not yet reproduced at scale; larger validation is ongoing. The library ships 186 property-based test functions across 13 files (written, not asserted here as all-passing).

That is the whole honest picture: a small, defensible core; encouraging measurements on synthetic and taxonomic data; and a flag planted firmly over a field we have not yet crossed. We would rather lose this bet in the open than hedge it into vagueness.

Common questions ​

What is a tensor-hyperbolic graph (THG)? A knowledge graph embedded in the Poincaré ball where each edge carries a tensor — an O(d) rotation derived from the two nodes' positions — rather than a single scalar weight. "Tensor-hyperbolic graph" is our own framing for this synthesis, not an established literature term, and the THG acronym is used elsewhere in the literature for unrelated work.

Is a tensor-hyperbolic graph a hypergraph? No — the two are easy to confuse but different. A hypergraph generalises the edge: one hyperedge joins many vertices at once. A hyperbolic graph keeps ordinary pairwise edges but places the graph in hyperbolic space. THG is the second kind — every edge connects exactly two nodes (it carries a rotation tensor instead of a number), and the whole graph lives in the Poincaré ball. Many-vertex edges in hyperbolic space — a hyperbolic hypergraph — are a separate research line; THG is not that.

Is AI memory really intrinsically hyperbolic? That is our bet, not a settled fact. Trees and hierarchies provably embed in hyperbolic space with very low distortion, and hierarchy is everywhere in knowledge — but whether arbitrary knowledge memory is intrinsically hyperbolic is an open hypothesis we are still testing.

Why not just use a normal hyperbolic embedding? A plain embedding records where nodes sit. The tensor edge additionally records the rotation between neighbours, which lets us compose around a cycle and measure tension (Tax) — detecting where a clean hierarchy breaks, which a scalar-weighted graph cannot express.

Is THG running in the production memory engine? No. The shipping engine projects Euclidean embeddings onto the manifold, which is approximate. Native hyperbolic embedding is the missing piece. THG and SLoD are research foundations, not the current production embedding path.

Edward Izgorodin — research note, last reviewed June 2026.