attractor.space
/papers/references/experiments
/notes/substrate
P001Frame-Semantic Graph Constructionframe-semanticsconstructionUpdated just now→R001TermsreferenceUpdated just now→N001Frame-type diversitydiversityframeUpdated just now→N002Coordination phasephasedynamicsUpdated just now→N003CouplingcouplingstructureUpdated just now→N004Divergence/convergence cyclecycledynamicsUpdated just now→E001ExperimentsvalidationUpdated just now→N005Fractal compositionstructureUpdated just now→N006FrameframestructureUpdated just now→N007Gini coefficientmetricUpdated just now→N008Hill diversitymetricdiversityUpdated just now→N009MembranesmembranestructureUpdated just now→R002MetricsinstrumentationUpdated just now→N010OscillatorsphasestructureUpdated just now→N011ResonanceresonanceretrievalUpdated just now→N012ScalescalestructureUpdated just now→N014SubstratestructureUpdated just now→N015Task-appropriate behaviorbehaviordiversityUpdated just now→
CC-BY-NC
lab notes on substrate dynamicsCC-BY-NC
N014·Julian Fleck·Last update2026/06/25
draftv1

Substrate

structure

A substrate is the shared medium a population of agents reads from and writes back into — their memory, context, retrieval, and tool state. AI is increasingly deployed this way: populations of agents over a common medium, not isolated chatbots. Its defining property is the write-back loop: each write feeds back into the medium, changing what can be retrieved on the next turn. That loop holds for any dynamic knowledge base — a single agent and a store already have it — but our interest is the second-order effects it produces as use accumulates, which a population sharing the medium makes both larger and harder to see.

Each retrieval strikes a different chord: an agent pulls a particular configuration of frames — a subgraph — out of the substrate; the next turn pulls another.

Multi-agent systems have a class of failures that are collective — properties of the population, not of any single agent: premature convergence on a narrow region before exploration is complete; a local error propagating as others retrieve it as evidence and amplify it; failure to converge at all. Two open questions follow: whether the substrate carries early, measurable signals of such failure — signals that would precede any single agent’s behavioural expression of it; and whether reversible interventions on the medium can restore healthy coordination without overriding the agents. The aim: oversight that scales with the size and depth of a deployment rather than being outpaced by it.

We borrow the term from ecology: the medium an organism lives in and grows on, soil or a reef or a culture in a dish. Knowledge in a knowledge substrate is treated the same way: as a living system, not a record. Frames and the agents over them grow, couple, and decay; regions flourish or wither; boundaries form and dissolve. The recurring vocabulary — membrane, coupling, resonance, decay — keeps that ecological sense, and is meant literally.

We build the substrate as a frame-graph: typed units that compose and nest, so the same structure reads across scales. Built this way, the store is an active medium, not a record — topology, vocabulary, and traversal behaviour all evolve with use. Its dynamics are the object of study. The rest of these notes concern what sustained use by a whole population does to such a medium.

N012ScalescalestructureUpdated just now→N010OscillatorsphasestructureUpdated just now→N009MembranesmembranestructureUpdated just now→N006FrameframestructureUpdated just now→N005Fractal compositionstructureUpdated just now→N003CouplingcouplingstructureUpdated just now→