M1
Analogy
Bound structures map cleanly across domains, so analogies are first-class retrievals.
Structured memory
Memory that scales with structure, not with context length.
Analogy, counterfactual, compositional binding — measured at the million-entry mark.
binding
g ⊗ e → p
grounding ⊗ entity → bound percept · the unit of recall
What we mean by structured memory
Memory that scales with structure, not with context length
We represent knowledge as bound structures — products of grounding and entity vectors — rather than flat token streams. Concepts compose, decompose, and recombine without being re-derived from context every time. Multi-hop chains stay tractable as collections grow because retrieval is a walk over structure, not a search over a buffer.
What the binding gives us
Three behaviours we get for free
M2
Counterfactual
The same machinery that retrieves an analogy can re-role and re-run the binding to evaluate a counterfactual.
M3
Multi-hop recall
Chains of two, five, or ten hops are walks over structure. Cost grows with structure, not with token length.
Scaling
P@5 stays at 1.00 as N goes to one million
We measure precision-at-5 on a held-out multi-hop retrieval task. The structured-memory line stays at 1.00 across three orders of magnitude; the byte-level baseline drifts down as the collection grows.
structured memory byte-level baseline (reference)
P@5 1.00
Multi-hop precision at one million entries
Held-out evaluation, no leakage between training and retrieval.
+0.36
Slot-factored over byte-level on relational binding
Held-out role-swap test, multi-seed.
8 / 8
Scientific gates closed in v0.3.0
Every release blocker was a measured test.
Slot vs byte
Role-swap is where structured binding earns its place
“Agent A names target B” is not the same proposition as “agent B names target A”. Slot-factored binding holds the role assignment; byte-level retrieval flattens it.
statement slot-factored byte-level Δ
"agent A names target B" 1.00 0.63 +0.37
"agent B names target A" (swapped) 1.00 0.60 +0.40
compositional new pair 0.99 0.69 +0.30
role-swap test, held-out, n = 5 seeds. multi-seed report in [[beyond-transformers]].
Scales with structure
Memory cost grows with the relational structure of what is stored, not with the length of the context window. That makes multi-hop reasoning tractable as collections grow into the millions of entries.
How we evaluate it
Every claim is measured with held-out tests and multi-seed error bars. Where it matters, we run the same task in a slot-factored memory and in a byte-level baseline so the lift is attributable, not assumed.