Agent memory and context: why your AI forgets, and what to do

The complaint arrives in the same words every time: it was great last week and now it has forgotten everything. Usually nothing broke. The system simply never had memory in the way you assumed.

Three different things people call memory

The context window

What the model can hold in mind for one task. Large, but finite, and it resets. This is working memory, not storage.

Retrieval

Fetching relevant documents at the moment of the task. This is what most "memory" features actually are, and it works well when the source of truth is written down somewhere.

Durable state

Facts the system carries across runs — your tone, your rules, what happened last month. This has to be deliberately designed. It does not emerge.

Why summarising loses the thing that mattered

Long chains compress context at each hand-off. Compression is lossy and the model choosing what to drop does not know which detail was load-bearing. The exception, the client's preference, the one constraint — those are exactly what a summary discards, because they look like detail.

What to ask a vendor

• Where is my business context stored, and can I read and edit it?
• Does each task see the original goal, or a summary of it?
• If I correct the output once, does the correction persist next month?

Structure helps more than model size here — see why multi-agent systems fail.

Keep reading

• Every agent platform we get compared to, side by side