Working with AI

How to keep AI project context for months

Short answer

Maintain a short current-state note, a decision history that includes reasons, clearly labelled hypotheses and observed outcomes. Do not reload the whole archive every time. Give the AI the context that matters for the task in front of you.

Dark stone milestones on a pale surface, crossed by one continuous amber thread, through zones of light and texture that suggest abstract seasons.

A project that lasts several months is not just a very long chat. It changes direction, accumulates trade-offs, abandons ideas, discovers constraints and produces outcomes that sometimes force you to revisit earlier decisions.

If you use AI throughout that journey, the best method is not to save every exchange. It is to preserve the structure of the project.

The most important document is the current state

Keep one short page that answers six questions:

  1. What is the objective now?
  2. Where are we?
  3. What has been decided?
  4. What is still a hypothesis?
  5. What is blocked?
  6. What happens next?

This page should stay short. It is a restart point, not the complete archive.

Date it. An undated project state is dangerous because an AI cannot always tell whether it is looking at the present or an old snapshot.

Preserve the reasons behind important decisions

A list of decisions is useful. A list of decisions with reasons is far more useful.

Compare:

We are using solution B.

with:

On September 14, B was selected because A created a proprietary dependency and C did not support the offline requirement. Revisit this decision if B cannot handle the expected load.

The second version lets an AI reuse the decision and also recognize when the conditions that justified it have changed.

Separate four kinds of information

Give important information an explicit status:

Fact: something verified at a given point in time.

Decision: a choice adopted to guide action.

Hypothesis: an explanation or possibility that still needs evidence.

Instruction: a rule to follow until it is replaced.

Many continuity failures happen when an old hypothesis returns later as if it were a fact.

Preserve outcomes, not only intentions

Projects learn. A decision becomes more valuable when you later record what it actually produced.

Example:

Decision: run all maintenance updates every Monday.

Three weeks later:

Observed outcome: two incidents were detected later than before. Return to two update windows per week.

That result turns a note into reusable experience.

Do not inject everything into every task

A six-month project may contain hundreds of relevant items. Feeding all of them into every AI request creates noise and increases the chance that old context will influence the answer.

For pricing work, retrieve product and commercial decisions. For a technical incident, retrieve architecture, similar incidents and security constraints. For brand work, you probably do not need the history of database migrations.

Continuity is not the amount of context. It is the quality of selection.

Create a maintenance ritual

Once a week or at each milestone:

  • refresh the current state;
  • close resolved questions;
  • mark replaced decisions as superseded;
  • attach outcomes to important decisions;
  • remove obvious noise;
  • check dates and provenance.

A few minutes of maintenance prevents a long-term memory from turning into a graveyard of outdated truths.

When several AI tools are involved

If you move from ChatGPT to Claude, from a coding agent to a general assistant, or from a local tool to a cloud service, do not transfer only the transcript. Transfer the current state, decisions, constraints and reasons.

That is what makes context genuinely portable.

A decision is only worth something if its reason survives: why preserve the reasoning behind a decision details the format. To give information a status, see the difference between a memory, a fact, a decision, a hypothesis and an instruction. To decide what to inject, giant prompt or targeted context offers a method. The first reflex when you change tools is described in how to move project context from one AI to another, and everyday practice in how to stop repeating the same information. A week-by-week example appears in using Nawame on a long-running project, and the documentation covers import and export.

An important limit

No memory system removes the need to verify critical information. An AI can retrieve a correct old decision and still apply it to a situation that has changed. Memory improves continuity. It does not turn the past into permanent truth.

Your next prompt should not have to contain your whole story.

Start building a memory your AI tools can find again when it matters.

Start free