We often talk about long-term memory as if it were the missing piece needed to create persistent AI companions.
Give an AI access to its previous conversations, store important events, retrieve relevant memories, and the system can appear to remember who it is and what happened before.
But there is a deeper architectural problem:
Is a collection of memories enough to preserve identity?
Consider a few situations.
The underlying model is replaced with a newer one.
The memory database is migrated to another system.
Part of the autobiographical history is corrupted or lost.
The AI is moved from one physical device or robot to another.
A retrieved memory conflicts with what actually happened.
The current context contains an incorrect reconstruction of the past.
In each case, simply saying “the AI has memory” doesn't tell us what should remain continuous.
This is particularly important for companion systems, where continuity is not just about retrieving facts about a user. The system may need to maintain continuity of its own history, relationships, decisions, motivations, experiences, and understanding of what is real.
That suggests a distinction between memory and persistent identity. Memory is something an agent can retrieve.
Persistent identity is something the system must maintain across changes to the model, context, runtime, and embodiment.
This also creates several architectural questions:
I've been working on an architecture I call PACC — Persistent Autonomous Companion Core to explore these questions.
The basic idea is to treat persistent identity as a system-level architecture, rather than as a feature of the language model or simply another memory database.
PACC separates concerns such as autobiographical memory, identity state, provenance and reality boundaries, relationship continuity, motivation, deliberation, runtime/model separation, persistence and recovery.
I'm sharing this because I suspect the problem will become increasingly important as AI systems move from chat interfaces into persistent agents and physical companion robots.
I'm interested in how other developers approach this problem.
If an AI's underlying model can be replaced without replacing the AI itself, where should that continuity actually live?