Substrate Dynamics: From Memory to Medium (Transcript)

19. July 2026
Membranes forming around co-active regions of a shared knowledge substrate
TalkTranscriptSubstrate DynamicsCoordinationMembranes

Transcript of my talk at Foresight’s Secure & Sovereign AI Workshop in Berlin on 19 July 2026.

Shared memory as a coordination surface

Agentic Stack

I’m Julian, one of the grantees at the Foresight Berlin node. Thank you to Foresight for bringing me into this. I’m also part of the Atlas Research Group, a research network exploring the Living Web and what the next iteration of the web could look like.

Our work on Dynamics follows a question: can the shared medium that agents use as a knowledge base also become a surface for coordination?

Across the agentic stack, there are many places to intervene to support collaboration and safety. Our focus is the shared knowledge store. The work touches the protocol layer because agents exchange information by writing back into that store, but our main interest lies in what happens to the medium as they use it. Where do agents leave their experiences, and what second-order effects arise as others encounter them?

Much of the work on memory infrastructure treats the knowledge base as a system of record. We are investigating whether it can also mediate coordination.

The knowledge substrate

Knowledge Substrate

We use the term knowledge substrate, borrowing a metaphor from biology and ecology: a medium that sustains and nourishes an organism.

In our definition, a knowledge substrate is the shared medium that a population of agents inhabits. Agents retrieve semantic units from it and write back into it. This write-back loop is central to our work, because each contribution changes the conditions under which the next agent acts.

Retrieval governs what an agent sees and therefore what it can act on. Together with its instructions and skills, access to particular regions of that knowledge landscape shapes its possible trajectories.

Cohesion and shared attention

Cohesion

We establish cohesion between agents and between the semantic units in the graphs we build. These are composable graphs, nested across several layers. We look at how multiple agents retrieve the content within them.

When two agents retrieve the same subset of information at the same time, we couple those agents. We do the same for the content. This is a form of Hebbian learning for both semantic units and agents, which we use as a proxy for shared attention: what agents are interested in and what they are working on.

Because the substrate is legible, we can also make an educated guess about the subject of that activity. In our case, the content is text. Observing the knowledge store gives us a view of shared activity that is independent of the particular model an agent runs.

From hierarchies to stigmergy

From Hierarchies to Stigmergy

Multi-agent orchestration is often described in terms of swarms. Yet many of the systems I see reproduce corporate hierarchies, with their own coordination problems.

We are investigating stigmergy: agents leave traces in a shared medium, other agents pick them up, and coordination develops through these local interactions over time.

That locality also makes stigmergic systems difficult to guide. Steering collective activity requires some way to constrain how agents move through the medium and what they encounter.

Membranes

Membranes: semipermeable boundaries

We work with membranes: temporarily formed boundaries around a set of agents and the content they are working with. Adjusting those boundaries changes what agents can access in the knowledge substrate, giving us a way to steer their exposure.

Resonance: reaching across boundaries

We also compute resonance between membranes to regulate how much information passes between them. The size of a boundary governs the region an agent can encounter; its permeability governs exchange across that boundary.

Two possible applications

Collective failure modes

One application concerns secure AI. A legible substrate lets us observe patterns in what agents are working on over time. An agent repeatedly retrieving the same material in a circular pattern, for example, may be stuck.

There are also failure modes specific to the collective setup: problems that arise when a population shares a medium and that may not appear when we examine agents individually.

Applications: Sovereign AI

The other application is one I am particularly curious about, if we can get there. We could use resonance between membranes to vary how much information actors share with one another. Privacy would become a gradient, with control over the depth of disclosure, rather than a binary choice between reading and not reading.

In this proposal, overlap in shared attention history and knowledge would help govern the detail exchanged between participants. That could give us a way to modulate the density of information crossing a membrane.

Multi-agent cooperation

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

In the breakout session, I would like to explore where such a system could lead, what it might enable, and how it could affect multi-agent cooperation.

The work involves several shifts: from individual to collective activity, and from top-down to bottom-up coordination. Because the substrate changes in real time, it also opens the possibility of moving from discrete, post-hoc measurements of traces toward a more continuous view of collective activity.