Don’t Start From Zero: How Dialogue Synthesis Preserves Decisions Between AI Chats

When every new AI chat starts from zero — you may need Dialogue Synthesis: a simple, freely available model that lets what you have built together with AI survive across chats, tools and people. Four files and two simple phases are all you need.

If you have been using AI agents for a while, you will surely recognise the situation. You have had a really good dialogue with ChatGPT, Copilot or one of the other AI platforms. Together you have landed on a strategy, edited a comprehensive document, made good decisions. The chat is sharp, your ideas have improved from being kneaded by someone with time and patience, the feeling of flow is at its peak.

Then you start a new chat the next day — and the AI starts from zero.

• The AI has forgotten what you agreed on in the previous dialogue, and you have to explain everything again.
• The answers in new chats bloat and become generic instead of being as sharp as in the original chat.
• Decisions you have made are quietly renegotiated in new sessions, because the agent has no idea that they have already been made.
• Knowledge the organisation has paid for in working hours evaporates when a chat is closed.

This is not a sign that you are using AI wrongly. It is a consequence of how today’s AI services work: every new chat starts with no memory of the previous one.

The problem can be solved. In this article I present Dialogue Synthesis — a simple, freely available model that lets decisions, rationale and lessons from your AI dialogues survive across chats, tools and people. It makes AI use both better and easier.

What is Dialogue Synthesis?

Dialogue Synthesis is a document model that captures important decisions, rationale and reusable lessons from an AI dialogue and makes them readable for the next session — or for a colleague, a new employee or another AI agent. The model consists of four Markdown files that are freely downloadable and used together with any AI service.

The name comes from what the model does: a synthesis of the dialogue — not a verbatim copy of the entire chat, but a distilled document with what actually matters: which choices were made, why, what they lead to and what the organisation can learn from them.

The model is developed by Portabla Media and published as an open package under the CC BY 4.0 licence — it is free to use, adapt and build on, including within organisations, with a simple credit to the source.

Two problems the model solves

Most people who have worked with AI in longer projects have run into two specific quality defects that arise between sessions.

1. Memory loss

The AI summarises or loses details as the context window fills up. What was settled yesterday is gone today. The result: duplicated work, invalid decisions and frustration.

2. Form inflation — when the answers bloat

This is perhaps the most deceptive problem with AI — and the one that is often most puzzling. When the AI only receives a description of the decisions — but not the actual work product — it is forced to guess the level of detail. And then something characteristic happens: the answer that yesterday was five sharp points becomes twenty-five generic ideas.

Technically, the AI has not “become worse”. It simply lacks the material that makes it equally precise.

Dialogue Synthesis solves both problems with the same underlying principle: the substance and the reasoning must always travel together. The work product (the document, the code, the plan) and the dialogue synthesis (the decisions, the rationale, the lessons) are uploaded together in every new session.

How the model works — four parts

Dialogue Synthesis builds on four parts:

1. The work product — the substance itself: the document, the code, the plan that the work has produced.
2. The dialogue synthesis — the meta-context: the decisions, the rationale, the lessons from the dialogue.
3. The ingestion prompt — the steering: a ready-made text that is pasted into the new chat, locking the frames, requiring the same stringency and pointing out today’s goal.
4. The template — the standards: the rule set that owns definitions, classification and status values.

The point is that the AI does the heavy lifting — you review and approve. The template contains ready-made instructions that let the AI agent itself analyse the dialogue, identify the decisions, classify them and formulate the synthesis. A prompt in the package gives the agent the entire work order.

The only thing you as a user need to know is the basic setup: which files to upload in which phase — and how to review the result.

The four parts are delivered as a package of four files:

The package is available on GitHub: https://github.com/PortablaMedia/dialogue-synthesis. Choose which language you want to use — there is a folder for Swedish and one for English. Download the four files — that is the entire installation.

How to use the model: step by step

The model is used in two phases. Neither requires technical knowledge — only that you upload the right files in the right chat.

Phase 1: Create a dialogue synthesis (at the end of a working dialogue)

When a working dialogue starts to feel valuable — or when you approach the end of a session — it is time to create a synthesis:

1. Upload the template in the chat where you are working.
2. Paste in Prompt 1 (it is found in the user guide, section 8). It asks the AI to analyse the dialogue and create a synthesis based on the template.
3. The AI does the work: identifies the decisions, the rationale, the assumptions and the reusable lessons — and writes the log.
4. Review the result. This is your most important step. The principle is AI proposes, a human reviews: read through, check that the decisions are correct and that nothing sensitive (passwords, personal data) has ended up in the log.
5. Save the log as a Markdown file, e.g. Dialogue_Synthesis_v1.md, together with the work product.

Phase 2: Continue in a new chat

When you continue the work in a new chat — the next day, next week or at a colleague’s:

1. Upload the work product (the document, the code, the plan).
2. Upload the dialogue synthesis that was created in phase 1.
3. Paste in the ingestion prompt and write what today’s goal is.

That is all. The AI now has both the substance and the reasoning, and can continue with the same precision as in the original chat.

Does the template need to be uploaded every time? No. The template is for phase 1 (and when uncertainty arises, or when a new synthesis is to be created in the continued dialogue). In phase 2, work product, synthesis and prompt are enough — the prompt already contains the locked frames.

When is it time to create a synthesis?

Create a dialogue synthesis when any of the following occurs:

• The dialogue has become too valuable to lose — you would be annoyed if the thread disappeared.
• The chat is starting to feel saturated or the AI agent is approaching its capacity.
• A session or phase ends and the next step requires a new chat.
• A milestone has been reached and you want to lock in the decisions.
• You need to change agent, tool or person in the work.

Conversely: not every chat needs a synthesis. One-off questions and short errands manage on their own. The rule of thumb is simple: document decisions that someone may need to revisit.

Three things you need — and three you don’t

This is the model’s most important advantage: the work moves from you to the AI.

You need:

1. To understand the basic setup (this article should be enough).
2. To download the files and use them in the right phase.
3. To review what the AI proposes before it is saved.

You don’t need:

1. To write the documentation yourself.
2. To learn the technology behind AI memory — those solutions with vector databases and RAG that you may have heard of, but never needed to understand.
3. To install anything — the files work in ChatGPT, Claude, Copilot and all other chat services.

The model also works across service boundaries: a dialogue synthesis created in ChatGPT can be loaded into Claude — the knowledge belongs to you, not the platform.

Save the files — that is where the memory is

One key to getting AI to remember is to save the files where you can easily access them. That is the whole point of Dialogue Synthesis.

Two habits are enough:

1. Save the synthesis as a file (for example Dialogue_Synthesis_v1.md) next to the work product. The next time they are needed, they are uploaded together — the work product and the synthesis belong together, that is the model’s fundamental principle.
2. Never overwrite an old version. Instead, create a new file (Dialogue_Synthesis_v2.md and so on) when something changes. If a decision has changed, mark the old one as superseded and point to the new one — just as the template shows. You get a history that shows what changed and why, without old decisions disappearing.

And one more thing: the file is made to be shared — between chats, tools and colleagues. Therefore never write passwords, API keys or personal data in a dialogue synthesis.

Get started

1. Download the package: Dialogue Synthesis on Github
2. Read the user guide, it’s a part of the package (about 15 minutes)
3. Choose a finished AI dialogue you’ve had and create your first synthesis — notice the difference in the next chat

Frequently asked questions

License

Dialogue Synthesis is developed by Magnus Nilsson, Portabla Media. The model is licensed under CC BY 4.0 and is available on GitHub: Dialogue Synthesis.

Foto

Kelly Sikkema, Unsplash.