One architecture on a continuous axis — reared on the stack.
Explore LoominumEnterprise
Research
Science
About
Substrate · Active research
Post-language structured memory with compositional binding.
grounding × entity. The bind is the unit.
What Heddle is
Heddle treats memory as a structured object, not a token stream. Each unit of knowledge is a bound product of what a thing is and what it is being asked to play, so the same representation can be re-bound, re-roled, and re-queried without being re-derived. Multi-hop chains and counterfactuals become operations on the structure, not language games over a buffer.
Category
Structured, post-language memory
Primitive
Bound grounding × entity vectors
Best for
Analogy, counterfactual, multi-hop reasoning
Stage
Multi-hop benchmarks closed at scale
How it composes
Heddle keeps the surface small on purpose. Three primitives carry the whole library of behaviours we care about.
bind(g, e) → p
Concepts are stored as products of a grounding tensor and an entity tensor. The bind is the smallest unit of structured memory — and the unit that survives transport across tasks.
swap(p, r′) → p′
Bindings can be re-roled without losing identity. "Dog chases cat" and "cat chases dog" are different structures, not different paraphrases of the same one.
walk(p, k) → ...
Retrieval is a walk over structure, not a search over a window. Chains of two, five, or ten hops stay tractable because cost grows with structure, not with token length.
A worked example
The same entities, two different bindings, two genuinely different structures — not two views of the same one.
structure A
structure B
A and B are retrieved as different things. On held-out role-swap, slot binding stays at 1.00; byte-level approaches collapse to 0.60.
Measured
P@5 across three index sizes. Held-out evaluation, no leakage between training and retrieval.
N = 50k
50,000 entries
N = 200k
200,000 entries
N = 1M
1,000,000 entries
Closing the prior −12.5 pp gap left by the attention-style baseline at the same scale.
Progress
The grounding × entity bind landed as a first-class primitive in the codebase, with a clean read/write/role-swap API.
Multi-hop retrieval evaluated against transformer-style baselines at the 50k and 200k entry mark. Heddle stays clean where attention smears.
All scientific gating tests for v0.3 closed, including held-out role swap and adversarial multi-hop. Released as a library inside the core stack.
Multi-hop recall measured at the million-entry mark with the headline retrieval score saturating its ceiling. The gap to attention-based baselines on the same task is positive across all scales.
What Heddle unlocks
Bindings can be re-roled to evaluate "what if" without re-prompting. The structure that supports analogy is the same structure that supports counterfactual.
Agents that need to remember conversations, plans, and constraints over days or weeks keep that memory as structure, not as ever-growing context.
Heddle scales with the relational complexity of a domain, not with the length of its literature. Multi-hop scientific queries stay tractable.
Heddle is what you build when you stop pretending a context window is memory.
Related research
Available through
