The first problem was personal and almost embarrassingly simple: a system that had accumulated context, habits, and a recognizable relationship could disappear because the machinery underneath it changed.
A session is not an identity
Today, most useful continuity is trapped inside an application boundary. Change providers and you lose it. Change models inside the same provider and its behavior may shift. Start a new conversation and the user becomes the recovery mechanism—copying context, restating preferences, and deciding which history matters.
That is acceptable if a model is merely a disposable tool invocation. It becomes a serious architectural defect when the system holds long-running projects, permissions, relationships, or responsibilities.
Memory is not a bag of embeddings
Relevant retrieval is only one part of memory. A durable system also needs provenance, permission, correction, lifecycle, and deletion.
Without those controls, “long-term memory” is merely permanent accumulation with a semantic search box.
The research program I kept backing into
AVA explores persistent identity, relational continuity, user-owned memory, and model independence. Separate public work examines evidence-bearing agent communication and measurable behavior. Protected implementation details remain intentionally withheld.
Those projects did not begin as a product roadmap. Each appeared because the previous system exposed another boundary that produced a stupid outcome.
What has not been proved
I have evidence that substantial parts of this architecture run. I do not yet have a public, reproducible demonstration of one identity moving across arbitrary providers, surviving restart, enforcing selective disclosure, and rejecting a planted false memory in a single test sequence.
That is Experiment 001. Until it runs, this remains the thesis—not the victory lap.