Fastest fix (as of June 2026): create a new Project with project-only memory turned on. In the new-Project dialog open More options and change Memory from Default to Project-only. From then on that Project draws context only from its own chats, files, and instructions, and your account-wide saved memories and past chats no longer bleed in. The catch: it can only be set at creation, not added to an existing Project. The rest of this page covers why the leak happens and how to fix Projects you can’t recreate.
ChatGPT’s “memory” now has four independent sources, and the UI mashes them together:
- Reference saved memories — the editable list at
Settings → Personalization → Memory → Manage(facts ChatGPT chose to keep or you asked it to). - Reference chat history — implicit recall of patterns from your past conversations, on by default for Free/Plus/Pro/Team. This one is new since the article most readers remember, and it is the most common surprise leak.
- Project Instructions / Files — scoped to the current Project only.
- The current chat’s own message history.
You expect “Project A stuff stays in Project A,” but global memory and reference chat history cross Project boundaries unless you stop them. Almost every “leak” is a layer-confusion problem.
Common causes
Ordered by hit rate, highest first.
1. Reference chat history is pulling from old conversations
The biggest change since this feature shipped. With Reference chat history on (default for Free/Plus/Pro/Team), ChatGPT recalls patterns and details from past chats account-wide, including ones outside the current Project. You ask a question in a JS Project and it answers like your last Python project because it remembered your habit, not because anything is wrong with the Project.
How to spot it: Settings → Personalization → Memory, toggle Reference chat history off, open a fresh chat, rerun the prompt. If the stale context vanishes, this was it.
2. Saved memories carry cross-Project preferences
You once said “I prefer Python,” ChatGPT saved it to your memory list. Now in a JS Project you ask “how to implement X” and it gives Python. A saved memory crossed the Project boundary.
How to spot it: Settings → Personalization → Memory → Manage. Entries unrelated to the current Project = this is a cause. (The saved-memory store holds roughly 1,200 to 1,400 words total as of June 2026, so old entries linger until you delete them.)
3. Project Instructions written as too-fuzzy / story-like
Instructions written as “this Project is about our e-commerce platform, we do SEO optimization” — ChatGPT uses that as background context perpetually, even when you ask unrelated questions and it still answers with an e-commerce tone.
How to spot it: Instructions over 500 characters, large blocks of “background / team description / project history” = this case.
4. Plus/Pro Projects reference other chats in the same Project (by design)
This corrects an older belief that “each chat in a Project is fully independent.” For Plus and Pro users, ChatGPT can reference previous chats within the same Project to stay focused. So a Project chat genuinely “remembering” an earlier Project chat is the intended behavior, not a bug. It only feels like a leak when two unrelated tasks live in one Project.
How to spot it: Re-run the same prompt in a totally unrelated Project; if the “memory” disappears, it was same-Project cross-chat recall, not an account-wide leak.
5. You leaked the context earlier in the same chat
In a 20-turn conversation you mentioned “the client said X” at turn 5. By turn 18 you forgot but ChatGPT is still pulling from the same chat’s context — not a leak, just an honest citation.
How to spot it: Search the current chat history; the info is actually there = legitimate reference, not leak.
6. Custom Instructions (account-level) affect every Project
Settings → Personalization → Custom Instructions applies to every Project. “I’m an engineer” makes every Project lean technical.
How to spot it: Temporarily clear Custom Instructions and rerun — tone / preferences change = this is interfering.
7. Team / Enterprise workspace settings override personal memory
In a Team or Enterprise workspace, admin-set workspace-level instructions may override your personal memory, but you assume it’s your memory leaking. (Memory is off by default for Enterprise/Edu and on by default elsewhere, which also changes behavior between workspaces.)
How to spot it: Run the same prompt in your Personal workspace — different behavior = workspace settings are the cause.
Which bucket are you in
| Symptom | Most likely cause | Go to |
|---|---|---|
| Stale habit/preference shows up everywhere, even brand-new Projects | Reference chat history (cause 1) | Step 1 |
| One specific fact (“uses React”) follows you across Projects | A saved memory (cause 2) | Step 2 |
| Project always answers in one tone regardless of question | Story-like Project Instructions (cause 3) | Step 3 |
| Within one Project, it recalls an earlier chat | Plus/Pro same-Project recall (cause 4) — usually intended | Step 4 |
| The “leaked” fact is earlier in this exact chat | Self-citation (cause 5) — not a leak | n/a |
| Behavior differs between Personal and Team workspace | Workspace override (cause 7) | Step 6 |
Before you start
- Confirm this is in Projects (not a Custom GPT or a plain chat).
- Duplicate the chat before retesting so history doesn’t pollute the next diagnostic.
- Confirm plan + workspace: Free/Go/Plus/Pro and Personal vs Team / Enterprise behave differently around memory.
Info to collect
- Project name + Project Instructions text screenshot + character count.
- Current
Settings → Personalization → Memory → Manageentries screenshot (redact PII). - Whether Reference chat history is on or off.
- Custom Instructions text (screenshot).
- A concrete “leak” example: you expected it not to know X, it stated X.
- Current model + Personal / Team workspace.
Shortest fix path
Ordered by ROI. The first two steps solve most account-wide leaks.
Step 1: Turn off (or audit) Reference chat history
If the leaked content is a habit or preference rather than one saved fact, this is the lever:
Settings → Personalization → Memory
→ Reference chat history → toggle OFF
→ Open a fresh chat, rerun the same prompt
→ Stale context gone = chat history was the sourceYou can leave it off for sensitive work, or leave it on and rely on project-only memory (Step 7) to wall a Project off from it.
Step 2: Prune saved memories of entries unrelated to the current Project
Settings → Personalization → Memory → Manage
→ Walk through every entry → Delete:
- Old-project preferences ("that project used React")
- One-off scenario preferences ("last time fixing a bug, used Python")
- Personal info irrelevant to current work
Keep:
- Always-true cross-scenario preferences ("reply concisely")You can also delete conversationally: tell ChatGPT Forget that I prefer Python and it removes that entry and confirms. Audit the list monthly; it caps around 1,200 to 1,400 words.
Step 3: Rewrite Project Instructions as “hard constraints + don’ts,” not stories
Bad to good:
Bad (800-char backstory):
This Project is about our e-commerce platform built in 2024 by the
marketing team. We sell organic products to women aged 25-40. Our
brand voice is warm and friendly. The team prefers Python because
historically we've used Django...
Good (150-char hard rules):
Constraints (always follow):
- Reply in English
- Output format: one paragraph, at most 300 words
- Tech stack: TypeScript only (no Python suggestions)
- Don't reference external products / competitors
Context: e-commerce SEO project for organic goods.Short and hard beats long and soft.
Step 4: One Project = one clear goal
If a Project mixes “SEO + customer service scripts + data analysis,” confusion is near-inevitable, and on Plus/Pro the same-Project chat recall (cause 4) will surface the wrong task’s context. Split into three:
- Project A: SEO
- Project B: Customer service
- Project C: Data analysis
Each with tight Instructions and dedicated files.
Step 5: Tighten Custom Instructions
If your Custom Instructions are too broad (“I’m a PM, like concise”), they override per-Project rules. Change to:
Default style preferences (overridable by per-Project rules):
- Concise unless asked for depth
- English by defaultThe explicit “overridable” tag helps ChatGPT let Project Instructions win when they conflict.
Step 6: Use Temporary Chat for sensitive tasks
Temporary Chat (top-right toggle in the model picker) by design doesn’t read memory, doesn’t write memory, and doesn’t keep history. Customer-private info, one-off research, anything you don’t want to influence future answers → always Temporary Chat.
Step 7: The real fix — recreate the Project with project-only memory
If a Project keeps absorbing outside context, the durable fix is to wall it off:
New project → name it → More options
→ Memory: change Default to Project-only
→ Recreate your Instructions + re-upload FilesWith project-only memory, ChatGPT references only the chats, files, and instructions inside that Project, and ignores your account-wide saved memories, reference chat history, and other Projects. Notes as of June 2026:
- It can only be chosen at creation. You cannot convert an existing Project, which is why this step recreates it.
- Web and the Windows app first; iOS and Android are rolling out.
- Sharing a Project automatically switches it to project-only memory, and that cannot be reverted to Default.
How to confirm the fix
- After pruning memory or turning off chat history, open a fresh chat and rerun the same prompt — leaked content gone = truly fixed.
- Run the same prompt in a totally unrelated Project — the current Project’s specifics don’t carry over = isolation works.
- For a project-only Project, confirm a fact you know lives in your account memory does NOT appear — that proves the wall is up.
If still broken
- Cut the Project to minimum: delete all Instructions + Files, keep one rule line, rerun — see if configs are fighting.
- Switch workspace: Personal → Team, run the same prompt, see if workspace settings are leaking.
- Fully rebuild as a project-only Project: create new, paste trimmed Instructions, don’t import old chat history.
- Package the saved-memory list + Project Instructions + a concrete leak screenshot and file a ticket at help.openai.com.
Prevention
- Audit saved memories monthly; delete expired / one-off preferences.
- For sensitive or self-contained work, start the Project with project-only memory from day one.
- Project Instructions = hard constraints + don’ts, never backstory.
- One Project = one clear goal; never mix 3 themes.
- Sensitive / one-off scenarios always go in Temporary Chat — never pollute memory or a Project.
FAQ
What’s the difference between “Reference saved memories” and “Reference chat history”?
Saved memories is an explicit, editable list of facts at Settings → Personalization → Memory → Manage. Reference chat history is implicit — ChatGPT learns patterns from your past chats and they shift over time. They have separate toggles; turn either off independently.
Why does ChatGPT remember things across my Projects even after I deleted the memory? Most likely Reference chat history (cause 1), not the saved-memory list. Deleting a saved memory doesn’t stop chat-history recall — turn that toggle off too, or use a project-only Project.
Can I add project-only memory to an existing Project? No. As of June 2026 it can only be set when you create a new Project. Recreate the Project and re-add Instructions and Files.
Do chats inside one Project share context now? On Plus and Pro, yes — ChatGPT can reference earlier chats within the same Project on purpose. That’s a feature, not a leak. Keep one Project to one goal so the right context surfaces.
Does turning off memory delete my saved memories?
No. Toggling Reference saved memories off just stops ChatGPT from using them; the list stays until you delete entries in Manage.
Related reading
- Project instructions ignored
- Project vs chat context
- ChatGPT memory not working
- ChatGPT project files not referenced
Tags: #ChatGPT #Troubleshooting