Working with files, project instructions and memory
Three ways to give a model persistent context, each with a different lifespan and a different failure mode.
What you will be able to do
Choose the right persistence mechanism for a given piece of context.
Before you start
- An assistant that supports file upload or project instructions.
Three lifespans
Attachments live for the conversation. Best for source material specific to this task.
Project or custom instructions live across conversations in a workspace. Best for stable things: your role, your audience, your house style, your standing constraints.
Memory features, where offered, persist across sessions and accumulate. Convenient, and the one to watch — because you stop knowing what is in it.
The stale-instruction problem
Standing instructions are excellent until they are wrong. A style rule you set in March quietly shapes every output in September. Review them on a schedule, and keep them short enough to read in one go.
The same applies more strongly to memory. If a tool has accumulated assumptions about you, know how to inspect and clear them.
What belongs where
Stable and about you: instructions. Specific and about this task: attachment. Anything sensitive: neither, unless the data policy for your account genuinely covers it.
Step by step
- List your standing instructions and read them as if new.
- Delete anything no longer true.
- Move task-specific material out of instructions and into attachments.
- Check what any memory feature holds about you and clear what should not be there.
- Put a reminder in the calendar to repeat this quarterly.
What good looks like
- You can recite your standing instructions.
- Nothing sensitive is stored in a persistent layer.
- Task material arrives as attachments, not as pasted history.
What it cannot do
- Memory implementations differ between products and change.
- File parsing quality varies, particularly for scanned PDFs and complex tables.
When it goes wrong
- Outputs carry an assumption you did not make today.
- Check standing instructions and memory. Something old is still speaking.
- Attachment appears ignored.
- Confirm it parsed. Ask for a direct quotation from it as a test.
Privacy and data handling
Persistent layers are the highest-risk place for sensitive data because they outlive your attention. Know how to inspect and delete them before you use them.
Sources3 referenced
- Claude developer documentation
Anthropic · official product documentation · link checked 2026-09-06
- OpenAI platform documentation
OpenAI · official product documentation · link checked 2026-09-06
- Guidance on AI and data protection
Information Commissioner's Office · regulatory guidance · link checked 2026-09-06
Links are checked at build time. If one has moved, tell us.
Read next
What context engineering actually means
Prompting is what you ask. Context engineering is what you put in the room before you ask it. The second is where most of the quality lives.
Checked 2026-09-06
Retrieval, source hierarchy and keeping context clean
Retrieval is how a model answers from documents it was never trained on. The quality of the answer is mostly the quality of what was retrieved.
Checked 2026-09-06
Privacy and sensitive information
The question is not whether these tools are safe. It is what you are permitted to put into them, and what happens to it afterwards.
Checked 2026-09-06