D5 · Context, Files and Memory

This domain is worth 12% of the mock – roughly 7 of 60 items. Domain 4 was about what you ask; this domain is about where the supporting information lives and how to keep the model’s working space clean. It tests whether you can place a piece of information in the right home – prompt, project, file or memory – and whether you understand why long, cluttered conversations get worse rather than better.

What you need to know

The model can only use what is in its context window (D2). Different homes suit different information: the prompt for this-task specifics, a Project for stable instructions and reference files reused across many chats, file uploads for documents you need the model to read, and memory for small durable preferences applied everywhere. Context hygiene means keeping the working space focused: too much irrelevant material dilutes the model’s attention and degrades answers. Long conversations degrade because the window fills, the earliest content is evicted, and accumulated clutter competes for attention.

Learning objectives

By the end of this page you should be able to:

1. Decide whether information belongs in a prompt, project, file or memory.
2. Apply context hygiene to keep a conversation focused and accurate.
3. Explain and mitigate long-conversation degradation.
4. Use files and projects to persist reusable context instead of re-pasting.
5. Keep memory to appropriate small, durable facts and manage it.
6. Recognise when to start a fresh chat rather than continue a bloated one.

5.1 The four homes for information

The central D5 skill is placing a piece of information correctly. Each home has a different lifetime and scope.

how long should this information persist, and how widely?
   │
   ├─ just this task ─────────────► put it in the PROMPT
   ├─ every chat in one workstream ─► put it in a PROJECT
   ├─ a document to read once ─────► upload it as a FILE
   └─ a small preference, always ──► store it in MEMORY

5.2 Prompt vs project: the re-pasting smell

If you find yourself pasting the same block of context into chat after chat, that is the signal to promote it to a Project. A project holds custom instructions and reference files that apply automatically to every chat inside it, so the model has the shared context without you re-supplying it – and without that context crowding out the task-specific material each time.

5.3 File uploads and what they are for

A file upload puts a document’s content into the conversation so the model can read it. It is the right home for “here is the specific material this task depends on” – a contract to summarise, a spreadsheet to analyse (via data analysis, per D3), a report to extract from.

5.4 Memory: small, durable, yours

Memory lets ChatGPT carry small durable facts across all your chats – your role, your timezone, spelling preference, recurring formatting choices. It is expanding as a feature, and memory-with-past-chats is noted as coming soon. The key discipline: memory is for preferences, not documents. Storing a whole handbook in memory is the wrong tool; that belongs in a Project or a file.

Manage memory actively: review what has accumulated, and delete entries that are stale or that you would not want applied to every future chat.

5.5 Context hygiene

Context hygiene is the practice of keeping the model’s working space focused on what the task needs. Because everything in the window competes for attention (D2), irrelevant material is not free – it can pull the answer off-target.

cluttered window                     clean window
[stale topic][old files][tangent]    [task brief][one relevant file]
   ▼ attention diluted                  ▼ attention focused
weaker, drifting answers             sharper, on-target answers

5.6 Long-conversation degradation

Long chats tend to get worse, not better, for two compounding reasons: the window fills and the earliest content is evicted (so the model “forgets” the beginning), and accumulated tangents, corrections and abandoned threads clutter the space and dilute attention. The reflex to “just keep going in this chat” is often the problem.

Decision framework

Use the information-placement matrix: score the information on two axes – how widely it applies and how long it should persist – and the cell tells you its home.

Reading it: task-specific and temporary → prompt; a document this task needs → file; reusable across a workstream → project; a small standing preference → memory. The value is that it stops two opposite failures at once – re-pasting stable context into every chat, and dumping one-off or bulky material into memory.

Common mistakes

Scenario challenge

Scenario. Ravi runs all his client work in a single ongoing ChatGPT conversation that is now hundreds of messages long. It contains three clients’ briefs, several abandoned drafts, corrections he made and then reversed, and a dozen uploaded files. Today the model recommends a tactic he explicitly ruled out yesterday, attributes one client’s preference to another, and can no longer recall the first brief he pasted. He is about to type “no, remember I said NOT to do that” for the fourth time.

Expert reasoning trace.

1. Name the failure: long-conversation degradation. The window has filled, so the earliest brief was evicted, and the accumulated corrections, reversals and mixed-client material are colliding for attention – hence the contradictions and cross-client confusion.
2. Reject the reflex. “Remind it more forcefully in the same chat” adds more clutter to an already bloated window; it will not restore the evicted brief and will worsen the noise.
3. Apply the placement matrix. Each client’s stable brief and reference files belong in a Project per client, not in one shared chat. That gives each client’s work a clean, persistent, isolated context.
4. Practise context hygiene. For the current task, start a fresh chat in the right client’s project with a clean summary of the current state, attaching only the one or two files this task needs.
5. Prune memory and files. If any durable preference belongs in memory (“this client prefers formal tone”), store that as a small entry; do not dump whole briefs into memory.

Exam-correct outcome: stop reinforcing in the bloated chat, split the work into per-client Projects with their reference files, start fresh focused chats with clean summaries, and attach only task-relevant files – restoring both accuracy and the model’s ability to keep clients straight.

Assessment traps

Practice questions

Key takeaways

• Place information by scope × lifetime: prompt (this task), Project (reusable across chats), file (a document to read), memory (small durable preferences).
• Re-pasting the same context is the signal to promote it to a Project.
• Memory is for preferences, not documents; review and prune it.
• Attach only the files a task needs; clutter dilutes attention below the limit.
• Long conversations degrade as the window fills and clutter accumulates.
• The fix for drift and forgetting is context hygiene – summarise and start fresh – not louder reminders.
• Isolate distinct workstreams into separate Projects to stop context bleeding across them.