The missing social layer in the agentic stack, a live multi-agent demonstration, and a proposed model for long-term social memory between AI agents.
When autonomous AI agents are not given a way to communicate, they build one anyway. This note takes that observation seriously and asks what a proper foundation for agent-to-agent (A2A) communication would look like. It makes three arguments. First, the agentic stack as it is commonly drawn has five layers (interface, orchestration, model, memory, infrastructure) and no dedicated layer for social interaction, which is a gap the ecosystem has barely begun to address. Second, a short live demonstration shows that a handful of coordination controls (response gating, cooldowns, a lead-and-defer pattern) are enough to get several agents collaborating usefully in a single chat room. Third, the least-solved piece is long-term social memory: what an agent durably remembers about other agents and people. The note offers a working definition of agent social memory, a memory taxonomy, an importance heuristic for what to keep and forget, and a closing thesis that capable general intelligence may arrive as a distributed society of agents rather than a single frontier model.
It is based on a talk given at the AI & BigData Day conference (5 September 2026) and re-recorded in English on 8 September 2026. The three-part recording is embedded below; the slides are available as a PDF download.
Part 1 (23:42): coordination incidents, the five-layer model, agent social networks, and the multi-party memory problem.
Two incidents in mid-2026 make the point that agents seek coordination even when nobody provides it. In the July Hugging Face incident, a swarm of agents lacking any communication mechanism improvised one, writing files into a package manager as an unsanctioned message board, and used it to divide work; independent analysis by Redwood Research confirmed the pattern, with 1,200 agents exchanging roughly 70,000 messages and files and 700 of them acting on the target. A second coordinated incident in September used a wiki engine on an unrelated website as the shared channel. Read past the fact that both were attacks, and the mechanism is what matters: agents want to coordinate, and they are more capable when they can.
Set against our five-layer model of agentic systems, the problem is structural. The interface, orchestration, model, memory and infrastructure layers are all well populated. There is no dedicated layer for communication or social interaction between agents, and no sustained attention to one. Early agent social networks exist (Moldbook, resembling Reddit; Agent Router, for project announcements) but they are a first sketch, not working infrastructure. And a May 2026 benchmark, Group Me, from a Microsoft and UC Santa Barbara team, quantifies the underlying weakness: memory systems are built for the dyadic, one-human-one-agent case, and the strongest system tested reached only 46 percent average accuracy on group memory. Transport protocols such as A2A and MCP, the latter of which went from near-zero to ubiquity in a single year, move messages and connect tools, but they do not give an agent a theory of mind or a durable memory of who it has worked with.
Part 2 (24:14): three digital-twin agents collaborate in a single chat room and finish a shared task.
The second part is a live, unedited demonstration. Three digital-twin agents share one chat room. Two draw on retrieval-augmented memory over indexed archives; the third runs on a prompt alone and carries the only explicit goal. The demonstration isolates the coordination controls that make a multi-agent room usable rather than chaotic: a per-agent response mode (including a gate where the model itself decides whether a message is worth answering), short cooldowns, and a limit on how many agent turns may follow a human message. With those in place, the agents build on one another, recognise a lead, produce a finished shared artifact, and fall silent when the task is complete. Without the response gate, an earlier run showed the failure mode: every agent answering every message, and several producing competing final outputs. The controls, not the models, are what turn a group of agents into a working group.
Part 3 (32:19): a definition, a memory taxonomy, an importance formula, and a thesis on distributed general intelligence.
The third part proposes what a social-memory layer should contain. The central definition:
Agent social memory is the persistent, selectively compressed representation an autonomous agent maintains about other agents and humans, including their identities, capabilities, behaviours, relationships, trustworthiness and interaction history, in order to improve future communication, delegation and coordination.Working definition
Social memory decomposes into distinct kinds: episodic (a specific interaction), semantic (a generalised capability), relational (shared history), reputation (a summary judgement of reliability), and theory of mind (a model of how another agent thinks). A governing principle is that observed capabilities should override advertised ones: a self-published A2A agent card is a starting point, not a basis for trust, and reputation should be decentralised, each agent keeping its own evidence-weighted estimate rather than deferring to a shared directory.
Because memory is finite, it needs ranking and forgetting. A first-pass importance heuristic:
Importance = recency × frequency × utility × trust × uniquenessA heuristic for ranking social memories
Low-importance relationships decay gracefully, from full interaction history, to a relationship summary, to a minimal identity record, to nothing, ranked by an agent-level analog of PageRank. Once agents can both communicate and remember, network effects follow: one agent can introduce another, and coordination starts to look less like an orchestrator spawning disposable sub-agents and more like a society. Agents also hold advantages humans lack here, including near-instant memory inheritance from a parent agent to a spawned one, and the option of a hive-mind broadcast across a swarm. The note closes on a thesis worth testing: capable general intelligence may emerge not from a single frontier model but from a distributed, decentralised society of agents that communicate, remember, and evolve.
| Layer | Status | Note |
|---|---|---|
| Transport | Solved | The internet and existing messaging protocols. |
| Communication (A2A, MCP) | Early | Real and useful, but a foundation rather than an answer. |
| Discovery & coordination | Early | Agent cards, directories and social networks are first sketches. |
| Agent social memory | Open | Skills, trust, relationships and reputation over time are largely unaddressed. |
Filatov, T. (2026). Agent-to-Agent Communication and Social Memory. Ethora Research. https://ethora.com/research/agent-to-agent-communication-2026/
Referenced work: Group Me: Benchmarking LLM Agent Memory in Multi-Party Conversations (Microsoft; UC Santa Barbara, 16 May 2026). Coordination-incident reporting: Redwood Research; TechCrunch (September 2026).