Skip to main content
SubStat — cover

Featured work

SubStat

in progress (testing)

An experimental memory store on a spiking neural network — population-coded concepts, capped fan-out, snapshot persistence.

  • Go
  • Spiking neural network
  • Population coding
  • Snapshot persistence
  • BrainCore

The premise

What if memory wasn’t rows and indexes, but a network that fires? SubStat is a research spike — prototyped in Claude Science and later folded into BrainCore — that stores and recalls information as activation patterns over a spiking neural network, not as records in a table.

How it works

  • Population coding — each concept is a distributed assembly of neurons, not a single address. Recall degrades gracefully: lose a few neurons and the memory is still there.
  • Recall = spreading activation — querying seeds an activation wave through the synapse graph; the concepts that light up are the answer.
  • Fan-out cap at creation time — the hard lesson. If you let synapses form freely, the graph explodes combinatorially as the store grows. SubStat caps fan-out when a synapse is created, so the network stays bounded no matter how much you feed it.
  • Snapshot persistence — the whole network is serialised to a snapshot and reloads intact after a restart. Synapses survive process death.

Why it matters

It’s a different answer to “what is memory for an AI agent?” — not Ctrl+F over embeddings, but an associative substrate where related ideas are literally wired together. The population-coding and spreading-activation ideas graduated into BrainCore’s recall path.

Status and numbers

Research / work-in-progress — folded into BrainCore. What made it real rather than a toy, in numbers:

  • Fan-out capped at 48 per neuron at creation → synapses stay bounded at ≤ N×48, never combinatorial.
  • Hub-heavy learn 5.4 s → 329 ms after the scaling overhaul; 184 concepts = 14K synapses (not millions); dashboard renders in 2–8 ms.
  • Persistence proven live: restart → “restored 7 concepts, 590 synapses” → answered a question without re-learning.
  • 17 SNN tests incl. an Export/Import round-trip; exercised against a LoCoMo eval battery (7 + ~24 probes across 9 categories, gpt-4o-mini judge).

Lessons

  • Biological plausibility is seductive and expensive. The win wasn’t “brains are cool” — it was the bounded fan-out that made a spiking store tractable.
  • Persistence is not optional. A memory that forgets everything on restart is a demo, not a store.