Memory Is an Architecture Decision
A good employee remembers the current conversation, knows company policy and may recall a customer's preferences. They do not remember every sensitive detail forever, treat an old conversation as current policy or share one customer's history with another. Agent memory needs the same distinctions. This seven-part series covers why stateless agents frustrate people, why replaying the transcript is not a memory strategy, the four kinds of memory, write policies, the difference between retrieval and recall, POPIA as a set of design questions, and how to test memory as behaviour. One fictional South African online retailer carries the examples. The POPIA material is general information, not legal advice.
7 articles
- Part 1Why the stateless agent keeps asking the same questionA customer gives an AI assistant her order number, asks a follow-up, and is asked for the order number again. Part 1 explains why that happens, why the first fix is a small piece of application state rather than 'memory', and what the rest of the series covers.8 min read
- Part 2Conversation history is not a memory strategyThe quickest way to make an assistant 'remember' is to resend the whole conversation on every turn. It works in a demo and becomes the most expensive, least predictable and least private form of memory in production. Part 2 shows the numbers and what to assemble instead.7 min read
- Part 3Four kinds of agent memory, and why they need different rules'Memory' covers at least four different things: the current task, what happened before, what we believe about a customer, and how work should be done. Part 3 shows where each should live, who may write it and how long it lasts, and why procedural memory is really versioned business policy.7 min read
- Part 4Memory write policies matter more than the databaseTeams starting on agent memory usually begin by choosing a database. The decision that shapes everything is earlier: what is allowed to be written, by whom, as what kind of claim, with what evidence and for how long. Part 4 sets out a write policy you can actually use.7 min read
- Part 5Retrieval is not recallWhen an agent retrieves a stored memory, it has found a candidate, not a fact it is entitled to use. Part 5 sets out the four checks every retrieved memory should pass (permission, freshness, relevance and conflict) and a precedence order for when sources disagree.7 min read
- Part 6Agent memory, POPIA and South African organisationsAn agent that remembers customers is processing personal information, and a memory store with no deletion policy is a liability waiting to be found. Part 6 turns POPIA's conditions into plain-language design questions for agent memory. General information, not legal advice.7 min read
- Part 7Evaluate agent memory as behaviourYou cannot evaluate memory by inspecting the database. What matters is what the system does: remembering the right thing, forgetting when it should, never leaking one customer to another and never inventing a past. Part 7 sets out eight behaviours, the scripted tests that check them and how to make them a release gate.7 min read
Working through any of this in your own organisation?
These articles come out of delivery work rather than theory. If the problems in this series look familiar, a short scoping conversation is usually the quickest way to work out what is worth doing first.
Talk to CloudNala