Memory & Context
Why a great prompt cannot include what you forgot to mention
Short answer
A prompt can only contain what you think to put in it. Old decisions, their reasons and abandoned options fade from your mind before they fade from the history of the project, so they never reach the request. A context memory makes them available again without you having to remember them.

You spent ten minutes on your request. The role is clear, so is the goal, the constraints are listed, and you even gave an example of the result you wanted. The answer comes back clean and sensible, and it suggests exactly what you ruled out eight months ago.
The prompt was not bad. It was complete for what you had in mind while writing it. The problem lies elsewhere: you cannot write into a request something you no longer remember.
What a good prompt does very well
A prompt is the text you give an AI to explain what you want. Well built, it sets a goal, a frame, an answer format, sometimes a tone. It is an excellent tool, and for an isolated task it is usually enough.
If you ask for a rewritten email, a summary of a document pasted into the conversation or an explanation of a concept, everything the AI needs fits in the request. No long history is involved. A good prompt helps the AI think, and it does not need to do more than that.
Immediate context and longitudinal context
Two kinds of context often get mixed together in a conversation.
Immediate context is what sits in front of you when you make the request: the task, the documents you paste, today’s constraints. It can be written down, and a good prompt writes it down well.
Longitudinal context is what happened before: the decisions made, the reasons, the options abandoned, the results observed, what replaced what. It can be written down too, but on one condition: that you think of it. And it is exactly what time erases first. A decision settled long ago no longer seems worth mentioning. An abandoned option does not come to mind when you are looking for options.
So the careful prompt handles the first kind of context perfectly and stays blind to the second, without you having done anything wrong. The temptation is then to paste everything into the request; what that costs is covered in giant prompt or targeted context.
An example: the trial nobody mentions anymore
Imagine Camille, who runs a small online shop. A year ago she tested a cheaper carrier for two months, then gave it up because too many parcels arrived damaged. She settled the question and moved on.
Today she asks her assistant: “How can I reduce my shipping costs?” Her request is excellent: volumes, margins, delivery times promised to customers. The answer recommends the cheaper carrier first.
On paper, that is good advice. In Camille’s history, it is a dead end already explored. She did not mention the trial, neither out of carelessness nor by choice: it was simply no longer in her head. No rewording of the request would change anything, because the missing information is precisely what Camille does not remember.
This is where history changes the analysis. With the past trial in front of her, the question becomes a different one: what failed last time, and is it still true? Maybe new packaging fixes the problem, maybe not. The decision stays with Camille. But she makes it with the right file.
What Nawame adds
Nawame is a context memory for humans and AI. It keeps the decisions, reasons, changes and results that give meaning to a story over time, then helps the AI tools you authorize find them again when it matters.
In Camille’s example, the carrier trial, the reason it was dropped and the date would have been kept. When the question about shipping costs comes up, the connected AI could report a precedent found: “you already tested this option, and here is why it was stopped.” It can also flag a possible connection, when the link is plausible without being certain, or admit no convincing precedent when nothing really resembles the question.
One point of vocabulary matters: found does not mean true. A retrieved memory arrives with its provenance and its date, so you can judge whether it still applies. Nawame provides context. Judgment and decisions stay human.
A good prompt helps the AI think. Nawame gives it a history to think with. For the difference between merely storing information and returning usable context, the article AI memory vs context lays out the reasoning. And if a decision has been replaced since, what happens to the old one has its own article.
When Nawame adds nothing
It would be dishonest to claim that a context memory is useful for everything. In several situations you do not need one.
- A one-off question. No history is involved, the prompt is enough.
- A project that fits on one page. If you reread that document before every session, it already does the job.
- Working with a single assistant. That tool’s native memory may suit you perfectly if you stay with it and what it retains is enough for you.
- Context you already pasted in. When the important facts are in the request, the AI sees them.
The gap appears when work is spread over time, with decisions piling up, sometimes across several tools, and a history that exceeds what you can paste in each time.
Current limits
Nawame cannot retrieve what was never kept. If a decision was recorded nowhere, it does not exist in the memory, and no tool will guess it.
Keeping too much is a flaw as well. Not everything deserves to become a lasting memory: not writing is a feature. Our first synthetic non-writing protocol correctly classified 26 events out of 26; the real world will be more complicated, with borderline cases that this protocol does not cover.
Finally, finding the right precedent remains a hard problem. A connection can look relevant without being so, noise can creep in, and old information can lose its value without anything signaling it. These are subjects that the Lab studies and measures, without claiming to have solved them.
What this changes for you
You do not have to become an archivist. A few habits are enough:
- Write your prompts for the present: the task, the constraints, the result you want.
- When you make a decision that matters, note its reason too, in one sentence.
- When you abandon an option, say why.
- Now and then, ask the AI: “have we already tried something close to this?”
The rest, the history that no longer fits in your head, can be handed to a memory that keeps it for you (a method without a dedicated tool is described in keeping AI project context for months). Your tools change. Your story stays.
If this sounds like your situation, the Who it’s for page describes where Nawame actually helps, and the pricing page explains what the free plan includes (up to 10 memories). You can also start free and try it on a single project.


