# D5 · Context, Files and Memory
What belongs in a prompt versus a project, a file or memory; context hygiene; and why long conversations degrade – with a decision framework for placing information.
import { Accordions, AccordionItem, Tabs, TabItem } from '@prosefly/astro-components';
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.
| Home | Scope | Lifetime | Put here |
| --- | --- | --- | --- |
| **Prompt** | This message | This turn | Task-specific details, the immediate question |
| **Project** | Every chat in the project | Until you change it | Reusable instructions, style guides, reference files |
| **File upload** | The chat you attach it to | That conversation | Documents to read, analyse or summarise |
| **Memory** | All chats | Until you edit/delete it | Small durable preferences about you |
```text
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
```
:::tip[Assessment signal]
"We re-paste the same guidelines into every chat" → a Project. "Remember I prefer British spelling" → memory. "Analyse this specific report" → a file upload. "For this one email, use a formal tone" → the prompt. The discriminator is scope × lifetime.
:::
## 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.
| Situation | Wrong home | Right home |
| --- | --- | --- |
| Brand voice pasted into every post | Prompt (each time) | Project instructions |
| A one-off tweak for a single email | Project | Prompt |
| Three reference PDFs used all week | Prompt (re-paste) | Project files |
| A document you will ask about once | Memory | File upload |
## 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.
Upload the actual document the task is about, ask a focused question, and verify the answer against the source. The file is grounded context the model can point to.
Uploading a dozen loosely related files 'just in case'. This clutters the window, dilutes attention, and can degrade answers even without hitting the limit. Attach only what the task needs.
## 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.
| Good memory entry | Bad memory entry |
| --- | --- |
| "I work in the EU timezone." | The full 40-page employee handbook. |
| "Prefer British spelling and metric units." | A one-off fact only relevant to one task. |
| "I lead the payments team." | Confidential data that should not persist. |
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.
| Hygiene habit | Why it helps |
| --- | --- |
| Attach only the files the task needs | Less competition for attention |
| Start a new chat for a new topic | Prevents stale context bleeding in |
| Summarise long back-and-forth into a clean brief | Compresses what matters, drops noise |
| Move stable context to a Project | Keeps each chat lean and reusable |
| Remove or ignore obsolete instructions | Stops contradictory steering |
```text
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.
| Symptom | Cause | Fix |
| --- | --- | --- |
| Forgot an instruction from the top | Earliest turns evicted | Re-state it, or move it to a Project |
| Answers drift off-topic | Accumulated clutter and tangents | Start a fresh chat with a clean brief |
| Contradicts an earlier correction | Conflicting instructions in context | Summarise the current state and restart |
| Slower, vaguer over time | Bloated window | Compress to essentials; new chat |
:::tip[Assessment signal]
"After a long session the model started forgetting / drifting / contradicting itself" → the fix is context hygiene: summarise the essentials and start a fresh chat (or move stable context to a Project), not "remind it more forcefully in the same chat".
:::
## 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.
| | Applies to one task | Applies across many chats |
| --- | --- | --- |
| **Short-lived** | Prompt | (rare – usually promote to a project or reconsider) |
| **Durable, small preference** | Prompt (if truly one-off) | Memory |
| **Durable, document/reference** | File upload | Project (instructions + files) |
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
| Mistake | Why it happens | What to do instead |
| --- | --- | --- |
| Re-pasting the same context every chat | Habit from single chats | Promote it to a Project |
| Storing documents in memory | Memory sounds like storage | Use files or a Project for documents |
| Attaching many files 'just in case' | Feels thorough | Attach only what the task needs |
| Continuing one endless chat for everything | Continuity feels efficient | Start fresh chats per topic; summarise state |
| Ignoring accumulated clutter | It is invisible | Practise context hygiene deliberately |
| Blaming the model for 'forgetting' | It feels like a bug | The window filled; re-supply or restart |
| Leaving stale memory entries | Never reviewed | Audit and delete outdated memories |
| Putting one-off facts in memory | Convenient in the moment | Keep one-offs in the prompt |
## 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
| Trap | Why it is tempting | The discriminator |
| --- | --- | --- |
| "Just remind it again in the same chat" | Continuity feels natural | A bloated window needs a fresh chat / summary, not more clutter |
| "Store the handbook in memory so it's always there" | Memory sounds like storage | Memory is for small preferences; documents go in Projects/files |
| "Attach every possibly-relevant file" | Feels thorough | Clutter dilutes attention; attach only what's needed |
| "A bigger window means one chat can hold everything" | More tokens feels safer | Attention still degrades with clutter; hygiene matters below the limit |
| "Re-paste the guidelines each time" | It works | A Project persists them without re-pasting |
| "The model forgot, so it's broken" | Feels like a bug | The earliest turns were evicted; re-supply or move to a Project |
## Practice questions
A. In memory as one long entry.
B. In a Project's custom instructions and reference files.
C. Re-pasted each time; that is correct.
D. In a single endless chat.
**Answer: B.** Reusable instructions and reference files across many chats belong in a Project, which applies them automatically. A misuses memory for bulk content. C is the smell being fixed. D degrades over time.
A. The full 30-page style guide.
B. 'I prefer British spelling and metric units.'
C. A confidential client contract.
D. A fact relevant only to today's single task.
**Answer: B.** Memory suits small, durable preferences applied across chats. A belongs in a Project/file. C should not persist. D is a one-off that belongs in the prompt.
A. Repeat the instruction more forcefully in the same chat.
B. Summarise the current state and start a fresh chat with only the needed context.
C. Raise the temperature.
D. Upload more files to remind it.
**Answer: B.** Long-conversation degradation is fixed by context hygiene – a clean fresh chat with essentials. A adds clutter. C is unrelated. D worsens the clutter.
A. Files are never read.
B. Everything in the window competes for the model's attention, so irrelevant material dilutes focus.
C. Files reset the knowledge cutoff.
D. Files disable memory.
**Answer: B.** Attention is shared across all context, so clutter pulls focus off the task even before the hard limit. A is false. C and D are unrelated effects.
A. Pasted into memory.
B. Uploaded as a file to the chat.
C. Described from memory.
D. Stored as a custom instruction.
**Answer: B.** A specific document the task depends on is a file upload. A misuses memory. C loses the actual content. D is for reusable instructions, not a one-off document.
A. Start a new chat when switching to an unrelated topic.
B. Attach only the files the current task needs.
C. Keep all topics in one endless conversation.
D. Upload every file you might ever use.
E. Never summarise long threads.
**Answer: A and B.** Fresh chats per topic and minimal relevant attachments keep the window focused. C, D and E all add clutter and dilute attention.
A. Memory.
B. A Project.
C. The prompt for that message.
D. A file upload.
**Answer: C.** Task-specific, short-lived detail belongs in the prompt. A and B are for durable/reusable context. D is for documents, not a one-line instruction.
A. Add a note at the top of the chat.
B. Create a separate Project per client with that client's briefs and files.
C. Increase reasoning effort.
D. Store all three briefs in memory.
**Answer: B.** Per-client Projects isolate each context so preferences do not bleed across. A does not prevent clutter-driven confusion. C does not fix mixed context. D dumps bulk into memory.
A. They always improve as they get longer.
B. They tend to degrade as the window fills and clutter accumulates, evicting early content and diluting attention.
C. They never lose earlier content.
D. Length has no effect on quality.
**Answer: B.** Filling the window evicts early turns and accumulated clutter dilutes attention, so quality tends to fall. A, C and D contradict how the context window works.
A. Memory is the right home for a small durable preference.
B. Memory should be reviewed and pruned of stale entries.
C. Memory is the right place to store large documents.
D. Memory only works within a single chat.
E. Memory cannot be edited or deleted.
**Answer: A and B.** Memory suits small durable preferences and should be actively managed. C misuses memory for bulk. D is false – memory applies across chats. E is false – entries can be edited/deleted.
A. The model is broken; reinstall the app.
B. The bloated window is degrading quality; compress to essentials and start a fresh chat.
C. The knowledge cutoff moved.
D. Non-determinism increased.
**Answer: B.** A bloated window degrades answers; the remedy is hygiene – compress and restart. A misdiagnoses. C and D do not explain gradual degradation from length.
A. Re-pasted into each prompt.
B. A Project (as a reference file with instructions).
C. Memory.
D. A brand-new upload every time.
**Answer: B.** Durable, document-level context reused across many chats maps to a Project. A and D repeat the re-pasting problem. C misuses memory for a document.
## 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.