Memory & Context
What should happen when a decision changes?
Short answer
When a decision changes, the old one should neither stay as it is nor disappear: it should be marked as replaced, with its date and reasons. Erasing loses the why, piling up leaves two contradictory truths. Replacing while keeping the history lets you understand the current state without picking the wrong reference.

In February, you write somewhere: “We’re staying downtown.” In August, you think the opposite, and you note it nowhere. When you ask an AI for help in October, which of the two sentences will it believe? That is the question every changing decision raises, and most memory tools do not answer it very well.
Why a changing decision complicates memory
Two notions help frame the problem. The current state is what is true today: “we are looking for a quiet place to live”. The history is the sequence of what was true before, with the reasons for each change: “we wanted the city center, then changed our minds for this reason”.
A document you keep updating mostly holds the current state, and the old version disappears under the new one. A conversation history holds the whole story, but flat: nothing says which sentence is the most recent or which one replaces the other. Each solves half of the problem. We have already described why a prompt cannot include what you forgot, and a past decision is part of that.
An example: staying downtown, or not
Imagine a couple, Inès and Paul, looking for a place to live. This example is invented for illustration.
In February, they decide to stay downtown, because Paul has no car and they do everything on foot. In August, after months of working from home, Inès writes: “calm and a real workspace now matter more”. They start looking at houses farther out. In October, they ask an AI which neighborhoods to visit.
Depending on what the memory did with February, the answer changes:
- It kept only February. The AI recommends downtown, confidently, and is wrong.
- It erased February. The AI suggests quiet neighborhoods, but cannot say why downtown was dropped. Nobody can check whether Paul’s constraint, having no car, was taken into account.
- It kept both, linked. The AI knows the current position is August, that February is replaced, and that its main reason (getting around without a car) may still need to be respected. It can ask the right question.
Erase, pile up or replace
Erase
It is clean and simple. You keep only what is true now. The cost is silent: you lose the why and the abandoned option. In six months, when someone asks “why didn’t we stay downtown?”, there will be nothing left to show.
Pile up
Everything is kept, in order. But two contradictory sentences side by side, with no status, leave the reader (human or AI) to guess which one holds. Nothing guarantees that it picks the most recent rather than the most detailed.
Replace while keeping the history
The current decision carries the status of reference. The old one is marked as replaced, with its date and its reason, and stays readable. You read the current state first, and the history only when it helps to understand. The record that keeps the reason with the decision is described in why preserve the reasoning behind a decision.
Why a false truth is worse than a gap
When a piece of information is missing, a well-run AI asks a question or says it does not know. When an outdated piece of information is presented as current, it answers with confidence, and the mistake goes unnoticed. That is why an unmarked old decision is more dangerous than an absent one.
There is also a more general trap: retrieved does not mean true. That a memory comes back proves neither that it is relevant nor that it is still valid. A good system must keep, with each memory, its provenance, its date and its context, so that you can judge (the checks to run before relying on a memory are detailed in why found does not mean true). We come back to this in the difference between memory and context.
What Nawame does with an old decision
Nawame is context memory for humans and AI. Among the things it can keep are precisely a change of position and what replaced an earlier decision, along with the reasons and the abandoned options. The replaced decision is not erased: it is replaced, and the history remains available.
Three practical consequences:
- Export lets you retrieve your memory with its replacements and provenance, not just the final state.
- In an organization, “authoritative context” means the reference status of a piece of information (a business rule, a policy, an official decision). When the policy changes, the new one becomes the reference. Authority is about the status of the information, not about it being impossible to delete. Organizations are described on the enterprise page.
- Imported information stays identified as imported and never becomes authoritative automatically.
Finally, Nawame provides context, and the decision stays human. Saying that February is indeed replaced by August is your judgment, not the memory’s. We develop this point in the article on retrieving past trade-offs without delegating.
When a simple document is enough
If decisions are rare and you work alone, a “Decisions” file with a date and a line reading “replaces the decision of…” does the job perfectly. If you only use one AI, on light personal topics, its native memory may be enough too.
The real cost is discipline: you have to remember to write the line at the moment of the change. A memory like Nawame becomes useful when changes are frequent, when several tools or several people are involved, or when you no longer remember yourself what was decided when.
Current limits
- Recognizing a replacement is not trivial. “We’re staying downtown”, then “actually, I’m looking for an office on the outskirts”: does the second sentence replace the first, or add to it? Sometimes only a human can decide.
- A replacement can be partial. A decision can change on one point and stay valid on the others.
- Replaced does not mean wrong. An old decision can remain informative, and one of its reasons can become relevant again. The other ways a memory goes stale are reviewed in how to tell that a memory has become obsolete.
- Nobody can guess a change that was never said. If the decision evolves in your head and not in the conversation, no memory can know. Nawame does not automatically know everything that is important.
- Time remains a hard problem. It is one of the topics the Lab studies and measures.
How to note a change right now, with any tool
When a decision changes, say so explicitly, with the reason. Three elements are enough: “New position: … Replaces: … Because …”. That short sentence helps a colleague, your future self and an AI alike, and it works in a document as well as in a memory.
If you want that history to follow you from one tool to another, you can open a first memory for free.


