Technology
A single idea runs through everything we build: meaning has structure. Retrieve on that structure — not on shared vocabulary — and you reach two problems keyword and vector search cannot solve: finding the same mechanism across domains, and giving agent fleets memory that is durable and governed.
Proof on real public data
We trained a structural bridge on ~250 example links between two incident records that share no vocabulary, format, or industry — NASA aviation safety reports and federal pipeline filings — then, for aviation incidents it had never seen, asked it to retrieve the pipeline accidents with the same underlying cause.
21.9%
Chance
38.4%
Text similarity
56.6%
Structural bridge
Retrieval accuracy is the reportable number. The working measure is agentic behaviour — whether an agent carrying this memory does the task better. That is what an internal agent gym scores, and what the research is steered by.
one pair · match 0.90
Read the two records first — then watch why every search you own would rank them as strangers, and what the engine read instead.
Aviation · Weather
Encountered unforecasted severe turbulence at FL250. No convective activity was present, and it was a relatively clear evening … the aircraft abruptly experienced large pitch and attitude changes … a continuous battle to keep the wings level.
Pipeline · Natural Force Damage
Water contained within the body of the valve expanded into ice as temperatures dropped; the expansion resulted in a hairline crack of the valve.
Two incident reports, filed years apart in different industries under different regulators.
Two accidents, not one word in common. The language never overlaps — so anything that matches on language is blind here.
Keyword and vector search rank by how much language two records share. With zero overlap, this true match scores like noise — the pipeline record never surfaces for the aviation query.
Strip the words away and the same three-step cause is left standing in both records. That skeleton — not the vocabulary — is what the engine retrieves on.
What the engine read
A latent force, unforecast
AVclear evening, no convective activity — turbulence that was never predicted
PLwater inside the valve, temperatures quietly dropping
acts on a system that read as normal
AVthe aircraft was flying fine at altitude
PLthe valve was intact and holding
and it fails, abruptly
AVlarge pitch changes — a battle to keep the wings level
PLthe expansion cracks the valve body
Clear air that turns violent; still water that turns to ice. Zero shared vocabulary — the match is the causal structure.
Three names, three layers
They are not interchangeable, and the distinction is the architecture.
Fractal Neurons
Holds the research and the discovery engine, and takes defined economics in the assets a partnership generates.
Space Neurons
A geometry-based retriever over a semantic vector space. Records are matched on the structure of what they describe, not on the words they happen to use — which is what recovers the aviation ↔ pipeline pairs above.
Dedicated Research Instance
Space Neurons running as the semantic operating system of a private, per-partner environment — a memory unit that spawns the agents working inside it, and improves as they work.
Why fractal
An instance can hold child instances. A machine that reasons over one partner's data is itself an agent inside a machine one level up, so the same architecture describes a single lab, a portfolio, and everything between. Drill into it and the shape repeats.
The agentic layer · associative memory
The substrate that surfaced those cross-domain pairs is the same memory we give to agent fleets. Agents store the experience they live, recall it by situation, and link experience to knowledge — and those links get stronger the more they earn their keep.
Decisions, findings, corrections, and outcomes are written to a per-agent store as they happen — passed through a novelty gate so repeated experience merges instead of piling up.
When an agent meets a situation it has been in before, the memory recognizes it and brings back what was actually used last time — personal experience layered over shared, immutable knowledge.
Every time a recalled memory is used, the link that surfaced it is reinforced; links that surface and get ignored weaken and fall away. Memory that improves with experience — never by overwriting what is shared.
Governed for fleets
Point that same associative memory at a body of knowledge instead of an agent's task history, and the links that make recall better become the correspondences no keyword search would surface. That is where the aviation ↔ pipeline pairs came from.
The deployment model · Dedicated Research Instance
We work with each partner inside a Dedicated Research Instance: a confidential, per-partner environment where their permitted internal material sits alongside selected external knowledge — bounded so that only they can reach it, with source boundaries preserved at every step.
Who we partner with
We work with a handful of partners in confidential technical domains. If you hold unique data with high discovery value, we'd like to hear what you're working on.
info@fractalneurons.com
