Guides
Using Nawame on a long-running project: a worked example over several weeks
Short answer
On a long project, you activate Nawame at the start, have it keep decisions, their reasons, ruled-out options and observed results, then pick the project up later with its anchor code. You can switch AI midway by connecting the new one to the same address. You check what was kept by asking what the AI retained, and sometimes the right call is to keep nothing.

This walkthrough follows one project from day one to week twelve, using a single made-up story. It shows the moves that matter: starting, deciding what deserves to be kept, resuming after a pause, switching AI, cleaning up. It uses only phrases and behaviors Nawame actually has, and it says where expectations should stop.
If you want the no-tool approach (a status file, a decision log), it is in the article on keeping project context for months. Here we assume you have decided to use Nawame and want to see how it goes.
The example project: an online shop launched solo
Meet Salome, a potter opening a small online shop for her ceramics, called Atelier Ronce. She works alone, part time, with an AI to think things through, write product pages and prepare her choices. The project will run for about three months, with quiet stretches in between.
What Salome does not want: to explain in every chat that she sells one-off pieces, that her budget is tight, that she already ruled out selling through a marketplace, and why. What she does want: for an AI, this one or another, to find those things when she needs them.
Week 1: connect and activate
First, the AI has to be able to talk to Nawame. Nawame is a remote MCP server: you paste a single address, https://hub.nawame.eu/api/mcp, into your AI’s connector settings. The account is created on first access, with a six-digit code sent by email. Nothing to install, no key to manage by hand. The exact steps depend on the AI: connecting ChatGPT to Nawame gives a detailed example, and the documentation covers Claude as well.
Once connected, Salome types this in her first exchange about the project:
active Nawame
(The phrase is quoted as Nawame gives it, in French. It opens a memory and starts saving as you go.) She then gives the AI the frame in ordinary words: the goal (open before autumn), the budget, the fact that the project is personal for now. She also keeps the anchor code of her topic within reach, something like NW-2ACRFY (an example code), so she can find it later.
Without that phrase, on a long-running piece of work, the AI may only do a light read at the start and suggest a grouped capture at a natural milestone. It creates nothing on its own.
What is worth keeping, and what is not
Over the first two weeks Salome settles several questions. Here is what is worth keeping, one example each:
- A decision and its reason. “I sell on my own site, not on a marketplace: I want to keep the relationship with my customers, even if the first sales come slower.”
- A constraint. “I cannot make more pieces than my kiln fires in a month.”
- A ruled-out option. “No international shipping at launch: breakage and customs, too risky for a first month.”
- An observed result. “Two friends tested the checkout: choosing the size of a piece was unclear.”
She also labels one thing as a hypothesis: “I think mugs sell better than bowls, but I have no data.” A hypothesis kept without its label turns into a fact three weeks later. The format for this is covered in why preserve the reasoning behind a decision.
What she does not have kept: the abandoned drafts of product pages, her dithering over the logo color, the small talk. A memory does not need everything. Not writing is a feature: sometimes the right memory action is to keep nothing.
At the end of a session, Salome asks:
qu’as-tu retenu ?
(Roughly: “what did you keep?”) The AI rereads what it has saved since the start of the exchange. Salome fixes one line, because the reason behind the shipping choice was worded badly, and moves on. This small check is worth doing every session.
Week 3: switching AI midway
In week three, Salome wants to try another assistant for writing her copy. Without an independent memory, this is where you paste in a rough summary by hand and lose the nuances.
Here she connects the new assistant to the same address, enters the code she receives by email, and types:
ouvre NW-2ACRFY
(“Open NW-2ACRFY.”) The topic resumes from its anchor code. The new assistant learned nothing from the other one: it reads the same memory, within what Nawame is allowed to return for Salome’s account. The AI never goes beyond the rights of the account.
One practical detail: at first Salome saw a shorter tool list than expected in the new assistant’s chat client. Some chat clients show a more limited list of tools than code-oriented clients do. She refreshed the connector and the tools appeared.
Week 9: picking up after three weeks away
Salome put the project aside for a craft fair. When she returns, she no longer remembers exactly what she decided about shipping costs, or why. She opens a chat and types:
ouvre NW-2ACRFY
Then, in plain words: “Tell me where the project stands, what is decided, what is ruled out and what was left to do.” The AI answers from the memory. Salome finds her decisions and their reasons in order, and above all the list of ruled-out options, which saves her from reopening a debate that was already closed.
She then checks: does what the AI says match what she knows? Found does not mean true: a line can be outdated, badly worded or pulled out of its context. The summary is a starting point, not proof.
Cleaning out what has become obsolete
At the fair, Salome learned that her market customers wanted delivery to a pickup point, an option she had ruled out. Her decision changes, and she simply tells the AI: the earlier position is replaced, here is the new one, here is why.
In Nawame, an item established as a reference can be corrected or replaced, and the memory can keep the fact that it was changed or replaced since, along with its provenance. The old decision does not vanish: it is marked as replaced, so that later you can still answer “why did we rule out the pickup point?” The general mechanism is described in what happens when a decision changes, and the signs of stale information in knowing when a memory has become obsolete.
A workable rhythm is a ten-minute tidy-up every two or three weeks. Asking “what is no longer true?” is enough.
Private, shared, and what you do not keep
By week twelve, Mael, a photographer, joins the project to shoot the shop’s pictures. Salome wants to share the calendar, the style constraints and the decisions about visuals with him, but not her notes on what she pays her suppliers, nor the doubt she has about one of them.
Nawame’s rule: personal memory is private by default, and what starts private does not leave private on its own. Sharing takes an explicit action, or a collaboration context with the right permissions. So Salome decides explicitly what becomes Shared. For the wider picture, why personal memory must not automatically become company memory develops the point.
And when she wants to talk through something without keeping it (a conversation about a difficult supplier, say), she types:
n’enregistre plus
(“Stop saving.”) The AI stops saving. Whatever was already written stays.
Current limits
This walkthrough describes what Nawame does today, not a frictionless experience.
- The example is invented. Salome, Atelier Ronce and Mael do not exist, and no result in this article should be read as real.
- You remain the judge. The AI chooses what to keep, and you reread it with “qu’as-tu retenu ?”. It can miss a decision or keep one too many.
- The phrases are quoted in French. They appear here as Nawame’s public material gives them. If you work in English, check the documentation for the wording your AI understands.
- Nothing works without a connection. Nawame requires an account, and it is called from the AI’s cloud, so your AI must be able to reach it. Disconnecting Nawame from an AI client deletes neither the account nor the memory.
- The tool list varies by client, as Salome saw in week 3.
- Stronger privacy has a price. A sovereign mode can reduce what the server is able to process, and so some features.
- A memory is a source, not an instruction. What is stored is data: the AI does not take it for an order.
For connection details, the Documentation page of the site stays the reference. And if you want to try this on a project of your own, Nawame is free for up to 10 memories: pick one topic, activate it, and come back to it in two weeks.


