What is agent memory and what it saves

Why this matters

• Understand why an assistant asks again about things it has already been told, and what that costs in operator hours
• Separate what is paid for in a knowledge base from what is paid for in memory: different work, different timelines
• Ask a vendor a question that can be verified: what goes into memory, for how long, and who can see it

What agent memory is in plain words

An electronics repair shop keeps a card on every client. A laptop came in during March, the keyboard was replaced, a discount on the next repair was promised. The intake clerk moved on; the card stayed. The new person opens it in ten seconds and picks up the conversation where it left off.

A language model works differently. Every call is its first. Past conversations settle nowhere on their own: everything the model "remembers" is sent to it again along with the current request.

Agent memory closes that gap. It is a store next to the model plus two rules: what to write down after a conversation and what to retrieve before an answer. The client card from the repair shop is exactly that.

How memory differs from the context window

The context window is a desk. The question, the instructions and everything needed for one answer are laid out on it. Once the answer is given, the desk is cleared.

Memory is the filing cabinet in the next room. It survives a restart, a change of model and yesterday. Before an answer, two or three relevant cards are pulled from it and placed on that same desk.

How it works

The fifth step is the one most often skipped. Writing is easy; unwriting is work that has to be designed in advance.

What goes into memory and what stays in the knowledge base

The split is simple: what is common to everyone lives in the knowledge base, what is specific and changing lives in memory. A delivery rate is the same for a thousand clients. A promise to call back on Thursday concerns one.

Mixing the two costs more. Common material is updated rarely, centrally and with review. Specific material appears every hour, goes stale within weeks, and has to be deletable at a person's request.

Hence a practical test. If a fact cannot be attached to a particular client, deal or task, memory is not where it belongs.

How this looks in a working process

A service company takes requests in writing. A client writes for the second time in a month: "about the same machine as last time." Without memory, the agent starts from scratch and asks for the serial number, the model and a description of the problem. The operator then deals with the irritation.

With memory, the same request pulls up three cards: the client's equipment, the previous request, how it ended. The answer becomes a continuation rather than a fresh questionnaire.

The benefit here is measurable, and it is worth measuring beforehand. Take one hundred consecutive requests over two weeks and label them by hand on two counts: the person repeats something already provided; the agent contradicts an earlier commitment. A month after memory goes live, the measurement is repeated on a comparable sample from the same process. Only samples of the same composition are worth comparing — season and request type shift the figure more than memory itself does.

What this gives the business

• A repeat request costs less than a first one. Part of the intake is not collected again — neither from the client nor from the operator. The saving is counted in minutes per request multiplied by the share of repeats.
• Commitments stop getting lost in handovers. A change of operator, a change of shift, a client returning a month later — the card stays put.
• The model bill stays under control. Three selected entries take up less room in a request than an entire message history. To get a sense of how much of the window is left for actual work, use the context headroom per conversation calculator.
• There is something to inspect. The contents of memory can be opened and read by eye. The agent's behaviour becomes explainable: it is clear which entry an answer came from.

When you can do without it

One-off tasks do not need memory. Processing an incoming invoice, classifying a request, extracting fields from a scan — everything needed arrives in the document itself, and the next document has no connection to the previous one. Such processes are launched without memory and lose nothing.

Starting without memory is also cheaper where the conversation is short and finishes in one pass. The answer itself is worth getting into working shape first. Memory is added as a second step — once measurement has shown that repeat requests make up a noticeable share of the flow.

What is worth checking

First: how long an entry lives. Every type of fact has its own span — a preferred contact method holds for a year, a request status stops meaning anything after a week. Retention is set before launch; otherwise, six months in, the agent will confidently cite a commitment that was cancelled.

Second: access boundaries. Memory holds information about people and deals. The same things are checked as with any database: where it sits, who can read it, how an entry is deleted at a person's request, whether the contents travel to an external service along with the request to the model.

Third: the selection rule. Ask to be shown, as a list, exactly what the agent decided to keep over the past week. The list takes ten minutes to read, and the skew shows up immediately: either everything is falling into memory, or nothing useful is. Without such a review there is no way to check the quality of memory.