Membranes
Every AI agent already works inside a boundary: its context — the instructions, memory, and retrieved material it can currently see. In a conventional setup that boundary is fixed: a user pastes in the prompt and reference material, and that constitutes the boundary of that agents working knowledge. This boundary begins to blur once an agent can explore it’s knowledge base of its own: the retrieval loops shape what it draws in on one turn and it’s output being persisted back to memory shapes what it might draw in next.
A membrane is that boundary treated as a first-class, dynamic object: selectively permeable, with channels that mediate information flow. The term is borrowed from cell biology, where a membrane both holds a cell together and regulates what passes across its edge.
Treating the boundary this way makes it two things: a place to read (the state of whatever works inside it is legible at the boundary, without inspecting each member) – and a place to act (adjusting its permeability changes what the enclosed agents have to work with, without steering any of them individually).
Over a knowledge substrate, those channels are gated by coupling density. A membrane in this definition is a temporary, semi-permeable boundary around a co-active region of the substrate: the collectively held frames, and the agents working over them, bound more tightly to each other than to their surroundings. It is drawn where coupling is dense, its channels open on a threshold (see resonance), and a decay mechanism releases it again.
Formation and dissolution
Membranes are established over the continuous evolution of the graph, not minted in discrete steps. Whenever context is retrieved from the substrate, the membrane is the region picked out by the current coupling density — the dense neighborhood the retrieval draws on that is set off against the sparser coupling around it.
What raises that coupling is co-attention — an agent reading, then writing, over a set of frames. It strengthens the coupling across the frames it draws together and sets the requisite variety inside the boundary: how much, and how varied, the context in play is (see also frame-type-diversity). The candidate rule we currently use is hebbian — strengthen coupling on co-retrieval events — but the underlying mechanism is still being worked out: frames carry an initial phase seeded from their content, and attention events drive them toward alignment from there.
The agent acts (closes a loop by writing traces back) and the next retrieval re-draws the boundary at the coupling density it then finds. Read and write are the same event class: both feed coupling. A membrane closes again as its frames fall dormant — between attention events the oscillators relax toward inactive, so the co-active region thins even as the underlying coupling persists. Without that relaxation a membrane would lock open and never reset.
Constituents
A membrane encloses both the information and the actors working with it, not one or the other (see fractal-composition). Nesting levels are not fixed: collapse a co-active region of frames to a point and it behaves as a single frame — a hole — though a live membrane is not quite a frame: it carries no declared schema, only an induced one (see frame). The reverse is also true: an agent in this sense is also a membrane – encompassing a bounded bundle of its harness state and currently active context.
Because we treat both actors and content as commensurable, we can apply the same mechanics at every level of nesting — activation, coupling, resonance, decay — and utilize the same metrics to read over a membrane as well as over atomic leaf frames.
Computing these metrics over subgraphs of frames is what picks out a membrane: two agents on the same codebase fall inside a frame naturally, and two people with overlapping research interests are grouped just as organically – close enough on the substrate to be put in proximity (or, in the human case, in touch – “you might want to talk”).
Nesting and overlap
A membrane is not a cluster. Clusters partition: each node lands in one group. Membranes nest and overlap — one can sit inside another, and two can share frames without either containing the other.
A membrane is held by both meaning and use, and the balance shifts over time. Each frame’s phase is seeded from its embedding, so a membrane starts semantically coherent — close in meaning, close in phase. Use then bends it: coupling built by co-attention pulls those phases toward synchrony (and a frame’s frequency drifts toward the ones it attends to), so frames drawn together by use can join the same membrane across semantically distant regions. Coherence is the starting point, not the definition — over time a membrane can come to span diverse regions, held by use more than by meaning, though retrieval still reads through embedding similarity, so it never leaves meaning behind entirely. An agent’s active context, a project’s working set, and a team’s shared region are membranes at different scales, drawn at the same moment over overlapping material. The object of study is membranes over membranes: boundaries layered and crossing, not tiles.
Selective permeability
A membrane is semipermeable: its channels open and close at thresholds, so flow across the boundary is conditional, set by what has accumulated on either side. What crosses should be gated on more than one dimension.
How deep a retrieval reaches is a threshold based on resonance — repeated actions on a set of frames and their semantic coherence. A high threshold returns only the most resonant frames, for example a summary at the top of a subgraph. When the threshold is lowered, retrieval reaches deeper and more granular detail is retrieved. This is progressive disclosure in the interface sense — the summary is always available; detail unfolds only as the threshold drops. Technically this is a traversal budget and a threshold setting in the retrieval re-ranker. The threshold steers granularity — how coarse or fine the retrieved context is (see fractal-composition).
Beyond depth, a membrane regulates two axes, and our substrate dynamics work is able to read it through either:
- Density. A membrane modulating the density of information flow through its channels — the requisite variety crossing in and out — and through that the capacity to act on that information. In this reading, disclosure policies and decision rights are emergent rather than primitive: the boundary’s permeability and the variety it admits produce an effective distribution of disclosure and authority, instead of that distribution being assigned agent by agent.
- Access. A membrane being semipermeable in the access-control sense — scoped, caveated, revertible exchange: who may pass which channel, granted in small reversible amounts. This is a reading further developed by Kenneth Bruskiewicz and Christina Bowen (Atlas Research Group) in what they call Mythic Membranes.
The two are complementary: one gates what and how much, the other gates who and how. The open question is which axis dominates for which problem, and whether they are two views of one gating mechanism or two mechanisms sharing a boundary.
Intervention and coordination
Deploying membranes over shared substrate is an active interest across a few projects, each coming at them from a different direction — from selectively permeable communication to trust-scoped coordination across teams. We mostly want a membrane to do two things: Mediate context availability — deciding what context a turn can reach, and (through resonance) at what granularity. And act as a coordination surface — agents that share a membrane coordinate through it: each one’s writes change the coupled region the others retrieve from next, so they align by acting on shared state rather than by messaging each other.
Because an agent is itself a membrane, the rule that builds any membrane builds agents: agents that fire together wire together. Co-acting agents accumulate coupling and are drawn together; the membrane around a team is a membrane over membranes, formed from the process of coordinating itself, bottom-up, rather than drawn around an org chart — and decay dissolves it again once the coordinating stops.
The membrane is also an intervention site. The variety it admits bounds what an agent can do: narrow context, limited reasoning and limited agency; broader inputs, broader reasoning and more room to act. So modulating the variety — the size and density of a membrane — is how we act on the agents inside it without touching them one by one. The same boundary makes their state legible from outside: read the membrane and you read what the agents have to work with.