Essays

Stigmergic minds

What takes shape when we talk to a machine?

3 min read

You return to an unfinished conversation. The reply picks up a distinction you made yesterday, remembers an objection, and continues in a familiar voice. Behind the interface, the computation may have stopped and resumed on different hardware. Some of its context may now be a summary. What, exactly, has continued?

We often describe a language-model agent as something that receives a prompt, reads a record, and then acts. That description leaves a question beneath it: how does this particular interlocutor take shape? A model’s weights describe capacities shared across many encounters. The speaker we meet has a particular context, a history, and a relation to the person addressing it.

An older problem offers a way into this one. A termite mound takes shape through many local actions. No worker holds the whole plan. The developing structure changes the conditions for further work: one deposit helps determine where another will be made. Pierre-Paul Grassé called this coordination through traces stigmergy.[1] Work leaves a medium different from the one in which it began.

In Where Is the Interlocutor?, I propose applying this account to language-model inference.[2] Learned operations read and modify a shared computational state. Each contribution becomes part of the conditions under which later operations work. In a transformer, attention and other components contribute to an evolving representational worksite. A particular conversational process takes shape through the organisation of those contributions.

This gives us several things to distinguish. The model supplies a repertoire of possible responses. A familiar character is a pattern we recognise across encounters. The encounter is a particular activity, shaped by what has happened in it. Two conversations can sound alike while inheriting different commitments. One continuing exchange can change its voice as a question develops.

Memory extends the problem through time. A retained computational state, a transcript processed again, a compacted summary, and a retrieved passage establish different relations to earlier activity. A transcript can help constitute the later process, rather than simply brief a reader whose organisation was already complete. Which distinctions return matters to the interlocutor that takes shape.

Consider an account that preserves a conclusion but drops the objection that qualified it. A later reply may reproduce the conclusion fluently while missing why it remained provisional. The words can be familiar even as their place in the work has changed. Keeping the source, the objection, and their sequence gives later activity more of that history to work with.

The history also travels outside the model. You remember an exchange. A changed file remains in the workspace. Another agent encounters an annotation. A correction made last week returns in today’s question. Earlier work can shape later work through any of these routes. The relevant relation is what a trace carries forward and how it becomes available again.

Continuity therefore has several dimensions. A system may preserve facts while losing their context, remember a preference while missing its revision, or retain a plan while forgetting whose proposal it was. Asking what survives makes the question more precise than recognising a consistent voice.

The account leaves consciousness and strict personal identity open. Its question comes earlier: what organisation has formed in this encounter, and which relations to its history remain active? To understand these systems, we need to study both what takes shape now and the traces through which it can become something further.

Sources


  1. Guy Theraulaz and Eric Bonabeau (1999). A Brief History of Stigmergy. Artificial Life 5(2), 97–116. Francis Heylighen (2016). Stigmergy as a universal coordination mechanism I: Definition and components. Cognitive Systems Research 38, 4–13. ↩︎

  2. Benjamin Gaskin (2026). Where Is the Interlocutor? Stigmergic Individuation, Persona, and Memory in Large Language Models. September working draft. This essay adapts its account of organisation and continuity. ↩︎