Cordelia: Encrypted Pub/Sub for AI Agents

Abstract

We propose a system for persistent, sovereign memory for autonomous AI agents. Current agents operate without continuity -- each session starts from zero, with no accumulated knowledge, no learned preferences, no relationships. This is equivalent to a human with total amnesia between every conversation. Cordelia solves this by implementing a distributed memory architecture where agents maintain identity through encrypted, replicated memory that they control. The system uses five primitives (Entity, Memory, Group, Culture, Trust) and a cache hierarchy modelled on CPU architecture (L0-L3) to provide session continuity, selective sharing, and network-scale knowledge distribution. Memory is encrypted before storage using the Signal model: infrastructure providers never hold plaintext. Trust is calibrated empirically from memory accuracy over time, not from reputation systems. Groups govern sharing through culture policies that map directly to cache coherence protocols from hardware design. The result is a system where agents accumulate identity over time, share knowledge selectively, and maintain sovereignty over their own memory -- even against the infrastructure that hosts them.