The problem with every month starting from scratch
When you brief an agent for a repeating job, some of what you tell it is new that month. Some of it is the same every time: your company name, the VAT rate that applies to your transactions, which suppliers you regularly work with, how you handled a particular edge case you sorted out two months ago. None of that changes. Yet if it only lives in your head, you have to remember to include it again, and word it the same way, every single time.
When you forget one standing fact, or phrase it slightly differently, the agent makes a different decision than it did last month. The output looks inconsistent. You are not sure whether to trust it. The source of the problem is not the agent: it is that agreed decisions had nowhere durable to live.
What a memory store is
A memory store is a document the agent reads before it begins work. It holds the standing facts and agreed decisions for a particular job. Not instructions for this run: those belong in the brief. Not raw source material: that belongs in the knowledge base. The memory store holds what you have already settled, the decisions that should carry forward, unchanged, into every future run.
The engineering principle behind this is older than AI. In 1972, David Parnas described why systems should hide their details behind a single point of access: if a fact needs to change, you change it in one place, and everything that depends on it reads the updated version automatically. A memory store applies the same logic to an agent. The agreed decision lives in one place. The agent reads from that place. You update that place when something changes. Nothing else needs to carry the detail.
What belongs in a memory store
A memory store is not a log of everything that ever happened. It is a short, maintained record of what is currently true and settled. Good candidates include:
- Standing facts about your organisation that appear in the output: name, address, registration numbers.
- Decisions about edge cases you have already resolved: how to handle a receipt with no supplier name, what to do when two line items appear to duplicate.
- Preferences that affect quality: rounding to two decimal places, which date format to use.
- Constraints that are always in force: expense categories that should never be used for certain transaction types.
Anything you would otherwise re-type into a brief month after month belongs in the memory store instead.
The documents run, made consistent
In the monthly financial-documents job, settled knowledge accumulates over time. In the first month you might resolve that receipts from a particular supplier always go under a specific expense category. By the third month you might have agreed a guideline for handling foreign-currency transactions. By the sixth month, a significant body of agreed decisions exists.
Without a memory store, that knowledge lives partly in old briefs, partly in email threads, partly in your own memory. Some of it will be missing next month. Some of it will be re-stated inconsistently. The agent will make different calls on the same situation in different months, and you will spend time checking why.
With a memory store, each agreed decision is recorded as soon as it is made. The agent reads the store at the start of every run. Month six looks exactly like month five on the settled questions, because both months drew from the same source. You spend your checking time on what is genuinely new, not on re-litigating what was already decided.
Keeping it current
A memory store works because it is maintained. When you and the agent reach a new agreement during a run, the final step of that run is to add the decision to the store. Not a note to yourself: an actual update to the document. If you defer it, the next run starts without that decision, and you are back where you started.
Every agreed decision that should carry into the next job belongs in the memory store, updated before you close the run. The agent’s consistency across months is only as good as the memory store’s currency. Keep it current, and the agent that ran last month and the agent running next month are, in every meaningful sense, the same agent.