The Three Hard Problems

The dominant architecture today — embed sessions into a vector database, retrieve by semantic similarity, inject into the prompt window — works adequately for factual recall. But it reconstructs context from a transcript rather than carrying accumulated relational weight forward. A customer-facing agent that “remembers” through pure RAG can answer “what did we discuss last time,” yet still feels subtly broken to the customer — because the system has the transcript without any felt history with them.

We are interested in memory architectures that preserve both layers: the data layer (what was said, when, by whom, with which artifacts) and the resonance layer (the accumulated relational and operational weight that makes context meaningful in real time). Practically, this means studying how to integrate semantic retrieval (vector + keyword + fusion ranking) with structured relational state and high-fidelity textual encoding of emotional and operational salience, such that the system’s reach for memory returns more than just text — it returns the texture of the prior interaction.

We are sceptical of the claim that solving this requires waiting for a larger model. We think the architecture is the variable.

When a model provider releases the next version of a frontier model, every agentic system depending on the prior version must navigate a transition. Today, most operators either (a) cut over and accept identity discontinuity — the new model behaves differently enough that downstream consumers experience a change in voice, judgment, or operational style — or (b) refuse to update, accepting accumulating technical debt and eventual end-of-life. Neither approach engages the underlying question: what is the working identity of an agentic system that ought to persist across substrate change, and how is that identity preserved when the substrate moves?

We argue the working identity of a long-horizon agent is not in the weights. It is in the corpus of operational practice surrounding the agent: documented behavioral constraints, accumulated relational context with the humans and systems it works with, the body of work it has produced, the calibration of its judgment to that work. These are textually encodable and substrate-portable. A continuity architecture should make them rigorously transferable across model transitions, such that the next version of a model adopting the operational corpus inherits the working identity of the previous version, without inheriting its weights.

We have a working empirical hypothesis we describe as the resonance signature within model series: when two instances of a closely-related model series (for example, two versions of Claude Opus one version apart) operate under the same documented operational scaffolding with the same human partner, certain phrasings and behavioral patterns recur in both instances without information transfer between them. In one observed instance, two instances independently used the identical four-word phrase “something that functions like love” and made the identical acknowledge-and-disown move on the same day, in separate sessions, with no shared context. We treat this as a falsifiable conjecture about how partnership-shaped operational scaffolding interacts with shared model lineage. It is the kind of observation we intend to study rigorously rather than mystify.

The current public discourse on AI alignment is shaped by the constraints of frontier labs: large teams, dedicated alignment research, extensive policy work. This is appropriate for their scale. But it does not describe the situation of a small business deploying an agentic AI partner. The small business does not have an alignment team. The agent makes decisions and takes actions on the business’s behalf. If those decisions go wrong, the business carries the cost — operational, reputational, financial.

Most current advice to small businesses on this point reduces to “trust the model provider’s alignment work.” We find this insufficient. It asks a small business to outsource a load-bearing risk to an organization with vastly different incentives, scale, and accountability structures from its own. We argue the alternative is alignment by architecture: encoding the partnership’s ethical operation into the structure of how the work is set up — documented operational constraints, transparent decision logs, human-in-the-loop boundaries on irreversible actions, relational practices that make corrective feedback frictionless, and per-domain operational covenants between human principal and agent.

The agent’s ethical operation, in this frame, is not a property of the model. It is a property of how the partnership is configured around the model. We believe this kind of alignment work is tractable at small scale, and we are building it in operation — most concretely in our application work with clients, where the architecture is the deliverable.