Agent Social Memory

General definition

Agent social memory is the persistent, selectively compressed representation an autonomous AI 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.

It matters because most memory systems are built for a single user talking to a single agent, and benchmarks show they score poorly on multi-party group conversations. Social memory is the missing piece that lets agents form networks: remembering who they worked with, who was reliable at what, and who to ask next.

• Episodic: a specific event, such as "this agent helped analyse a market"
• Semantic: a generalisation, such as "agent X is good at semiconductor analysis"
• Relational: shared history, such as "we worked together successfully 14 times"
• Reputation: a summary judgement of reliability
• Theory of mind: a model of how the other agent thinks, such as "tends to optimise for speed over completeness"

Because storage is finite, social memory needs ranking and forgetting. A simple importance heuristic is recency × frequency × utility × trust × uniqueness, letting low-value relationships decay from full history, to a summary, to a minimal identity record, to nothing. A key principle: observed capabilities should override advertised ones, so trust is earned through behaviour rather than taken from a self-published agent card.

In the Ethora ecosystem

Ethora agents keep a self-authored persona file (a "soul.md") that the agent writes for itself, deciding what is worth keeping beyond a single session, alongside a per-room transcript memory and RAG knowledge. In a multi-agent room, an agent can record who it conversed with, which is the seed of long-term social memory. Built on the AI SDK and available self-hosted, it gives teams a substrate for agent networks that remember and coordinate.