Substrate Dynamics: From Memory to Medium

19. July 2026
Membranes forming around co-active regions of a shared knowledge substrate
Substrate DynamicsMulti-Agent SystemsCoordinationMembranesSovereignty

What happens when a population of agents reads from and writes back into the same knowledge store? Each contribution changes what the next agent encounters. That loop could turn shared memory into a medium for coordination.

This article follows my talk at Foresight’s Secure & Sovereign AI Workshop in Berlin on 19 July 2026, in the track on decentralized and cooperative AI.

From memory to medium

Agentic Stack

There are many places to intervene in an agent system: model weights, prompts and skills, communication protocols, or the tools an agent uses. Our work on Dynamics focuses on the shared knowledge store.

Memory infrastructure is often treated as a system of record: somewhere to leave information for later retrieval. But a store that many agents read from and write back into also changes through use. Each contribution becomes part of the environment in which another agent acts. We are investigating whether this shared medium can support coordination as well as memory.

The knowledge substrate

Knowledge Substrate

We call this medium a knowledge substrate, borrowing the term from biology and ecology: a base on which an organism lives, grows, or obtains nourishment.

The defining property is the write-back loop. Agents retrieve semantic units from the substrate, work with them, and contribute new material. Those contributions alter what can be retrieved next. Retrieval therefore governs more than access to information: together with an agent’s instructions and skills, it shapes its possible trajectories.

Cohesion and shared attention

Cohesion

We establish couplings between agents that retrieve the same information at the same time, and between the semantic units they retrieve together. This Hebbian mechanism provides a proxy for shared attention: which agents are working with related material, and which parts of the substrate are active together.

Because the substrate contains legible text, these patterns can also tell us something about the subject of that activity. Observing the shared store offers a view of collective work that does not depend on inspecting the internal state of a particular model.

Coordination through traces

From Hierarchies to Stigmergy

Many multi-agent systems described as swarms still rely on hierarchical orchestration: a coordinator assigns tasks and passes messages down a control structure.

Stigmergy offers another model. Agents leave traces in a shared environment; other agents encounter those traces and act on them. Coordination develops through local interactions over time. The difficulty is how to guide such activity without introducing a central coordinator. Our approach is to constrain how agents move through the medium by shaping what they can retrieve.

Membranes and resonance

Membranes: semipermeable boundaries

are temporary boundaries around co-active regions of the substrate, enclosing both agents and the content they work with. Changing a membrane’s extent changes what its agents can encounter.

Resonance: reaching across boundaries

Resonance concerns the alignment between membranes. We evaluate the semantic units they share against their internal state to regulate exchange. The boundary’s extent governs access within a region; its permeability governs what passes between regions.

Secure and sovereign AI

Collective failure modes

For secure AI, a legible substrate could help reveal collective failure patterns. Repeated circular retrieval, for example, may indicate that agents are stuck. Some failures arise specifically from sharing a medium and may not be apparent when agents are examined individually.

Applications: Sovereign AI

For sovereign AI, selective permeability suggests a way to vary the depth of disclosure across a boundary. Participants might exchange fuller context, a summary, or nothing, rather than choosing only between public and private access. Using shared attention history and knowledge to govern that resolution remains a research direction.

Toward collective coordination

What could a legible, self-organizing substrate used by agents (and eventually people) become...?

This work shifts attention from individual agents to collective activity, and from top-down control to coordination through a shared medium. Because that medium changes during use, it may also support a more continuous view of collective activity than post-hoc inspection of isolated traces.

I’m taking this work forward through membranes.network, a protocol for selective, attributable context exchange between independently governed systems. The focus is on how systems can share useful context while retaining control over their boundaries, with provenance that lets participants trace what crosses between them.

Recording