Ask most teams how their AI agent will “remember,” and the answer is a vector database. Embed everything, retrieve by similarity, hope the right context comes back. For a consumer chatbot, fine. For an agent that touches money, it is the wrong default — and the reason is not accuracy, it is accountability.
An agent is not one big model
The first mistake is treating “the agent” as a single large language model. In practice an agent is right-sized intelligence behind a gateway: a business rule where the policy is known, a scoring model where the signal is statistical, a small language model for narrow domain tasks, a larger model only for the rare open-ended reasoning — every one of them proposing through a single control point that records what was proposed, by whom, and why. Most of the work lives at the cheap, deterministic end. That is what keeps an agent estate affordable and governable.
Memory is a projection, not a store
The second mistake is bolting on a separate memory system. The cleaner idea: the bank already has an ordered, append-only record of what happened — the event log. Memory is just a projection of that log. Rebuild a graph of relationships, a set of features, a similarity index — all derived from the one governed source, and all scoped by consent and authority so the agent can only remember what it is entitled to see.
This buys three things a vector store cannot:
- Replay. You can reconstruct exactly what the agent knew at the moment it acted — the question every auditor and every incident review asks first.
- Lineage by construction. The provenance is the log; you are not reverse-engineering it before an assessment.
- A flywheel. The same events that form the agent’s memory are the training substrate for models trained on the bank’s own behaviour — the proprietary history a competitor cannot replicate.
Why this matters now
As customers’ own agents start to switch deposits, route lending, and plan payments, the bank that wins is the one that can host inference in a governed slot — over data it controls, with a memory it can prove. A vector database gives you recall. A ledger-shaped memory gives you recall plus audit, replay, and consent. In a regulated institution, the second list is the one that ships.
We have proven this in a clean-room TypeScript pilot — the agent’s memory rebuilt live in the browser by replaying its own event log. The full architecture, the schemas, and the model that sits behind it are available to approved bank evaluation teams under a gated technical review. If that is you, get in touch.
