ChatGPT’s Memory is OpenAI’s “auto-profile” feature (launched 2024): facts you mention in passing (“I use macOS”, “I’m building a SaaS”, “I prefer short answers”) get written to a separate memory store and automatically injected into your system prompt next session. A few properties of this system are widely misunderstood: it’s not a verbatim transcript, it does not record everything, and in some modes (Temporary Chat / Custom GPT) it doesn’t work at all.
When it “doesn’t remember,” you’ve almost always hit one of these 6 cases. In rough order of frequency:
Common causes
1. Memory toggle is off / you’re inside a Custom GPT
Settings → Personalization → Memory is on by default, but some users turn it off for privacy. Custom GPTs don’t read or write your main-account memory by default — each GPT has its own instructions.
How to verify: Settings → Personalization → check Memory toggle; verify the current chat isn’t inside a Custom GPT (the GPT name shows in the top-left when you are).
2. You’re in a Temporary Chat
Temporary Chat by design doesn’t read or write memory and doesn’t store the conversation. It has a dashed border at the top + a “Temporary chat” label. Saying “remember this” inside one gets a verbal yes from the model, but nothing is stored.
How to verify: check for the “Temporary chat” label at the top.
3. What you said didn’t trigger a memory write
Not every sentence gets written. The model decides “is this a fact / preference / long-term info?” before writing. Vague tone, transient context, questions, jokes — none of these qualify.
| Won’t write | Will write |
|---|---|
| ”I’m a bit tired today" | "My working hours are 9–6" |
| "Let me look at this first" | "Remember, I build sites in Astro" |
| "Edit this paragraph" | "Remember, I want bilingual output” |
How to verify: start with “Remember that…” — when memory actually writes, the model’s reply shows a ”📌 Memory updated” badge. No badge = nothing written.
4. Memory store is full
The memory store has a limit (~100 entries, OpenAI tunes this). Once full, old entries may be auto-evicted, or new ones silently dropped.
How to verify: Settings → Personalization → Memory → count entries. Near the cap means clean up.
5. Memory was written / read in different workspaces
For Team / Enterprise accounts, Personal and Company workspaces have separate memory stores. Telling it “remember X” in Personal then switching to Company means it doesn’t know.
How to verify: the dropdown next to the avatar shows the active workspace. Both conversations must be in the same one.
6. Cross-device / cross-account sync delay
Memory sync to the app can lag by hours. What you just wrote on Web may not show up in the app yet.
How to verify: write something on Web, then immediately check the app’s memory list — missing = not yet synced.
Shortest path to fix
30-second checks first:
Step 1: Confirm Memory toggle + not in Temporary Chat
- Settings → Personalization → Memory → toggle ON
- Confirm current chat has no “Temporary chat” label at the top
- Confirm you’re in native ChatGPT, not a Custom GPT
Step 2: Manually review the memory list
Settings → Personalization → Memory → “Manage memories.” This is the canonical list of what ChatGPT actually remembers. Check:
- Is it in the list? What you thought it remembered — is there an entry? No entry = never written
- Conflicts? Two contradictory entries on the same topic (“prefers short” + “prefers detailed”) — model picks at random
- Stale? Old preferences may no longer apply
Delete conflicting and stale entries.
Step 3: Force a write with “Remember that…”
Test template:
Remember that I prefer responses under 200 words unless I ask for detail.After sending, look for 📌 / “Memory updated” near the bottom of the reply. Present = written. Missing = the model decided not to remember; reword as a more “long-term fact” statement.
Step 4: Free up space
If the list is at the cap:
- Bulk-delete entries you no longer need
- Re-tell the model your core profile in one go (“I am X, I use Y, I want Z”)
Step 5: Pin permanent things to Custom Instructions
For anything that absolutely must persist, use Settings → Personalization → Custom Instructions (“About you” + “How to respond”). These two fields are baked into every new conversation’s system prompt — more stable than memory.
Template:
About me:
- Work / project I'm doing
- Preferred answer style (length, tone, format)
- Common tool stack / platform
- Key constraints (don't do this)
How to respond:
- Conclusion first, then reasoning
- Code in Markdown blocks
- Keep English technical terms in original when mixing languagesStep 6: Switch to the right workspace + wait for sync
Team users: verify memory writes and reads happen in the same workspace.
App not synced: force-quit the app → relaunch → wait 1–2 minutes.
Easy to misdiagnose as
“It remembered me last week but not today” usually means you switched into Temporary Chat, sync broke during a cross-device hand-off, or your workspace switched.
Prevention
- Treat memory like a small notes file — review it monthly and remove stale / conflicting entries
- Don’t expect cross-account memory transfer — one account = one memory pool
- Pin truly important facts to Custom Instructions instead of relying only on memory; more stable, doesn’t eat memory quota
- Project-level context (one article’s style, one client’s preferences) belongs in Projects, not memory
- Don’t do work you want to keep inside Temporary Chat — check the top label before starting
Related
- ChatGPT projects guide
- ChatGPT history missing
- ChatGPT beginner guide
- ChatGPT prompt improvement
- ChatGPT model selection guide
Tags: #ChatGPT #Debug #Troubleshooting