ChatGPT Memory Not Working: Why It Forgets and How to Fix It

ChatGPT says "got it" then forgets next session? It's usually a toggle, Temporary Chat, a write that never happened, or the new background memory. Check these and fix it.

Published: May 17, 2026 Updated: Jun 15, 2026 Author: AI Productivity Guide Team 🌐 查看中文版本

Fastest fix: open Settings → Personalization → Memory, confirm both Reference saved memories and Reference chat history are ON, make sure the current chat has no Temporary chat label at the top, then click Manage memories and check whether the fact you expected is actually in the list. If it’s not there, it was never written — re-state it as a plain fact and watch for the Memory updated line in the reply.

ChatGPT’s Memory is OpenAI’s auto-profile system: 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 injected into your prompt in later sessions. Three things about it are widely misunderstood: it is not a verbatim transcript, it does not record everything, and in some modes (Temporary Chat, some Custom GPTs, Projects) it doesn’t touch your main memory at all.

As of June 2026 there are now two layers you have to think about, and a 2026 change made the behavior less obvious:

Reference saved memories — an explicit, editable list of facts ChatGPT chose to keep. This is what you see under Manage memories.
Reference chat history — implicit recall of patterns from past chats (launched April 10, 2025). There is no list for this layer; ChatGPT just draws on prior conversations.

On top of that, OpenAI began rolling out the “Dreaming” memory update (Dreaming V3) on June 4, 2026 (Plus and Pro in the US first; Free, Go, and international following in the weeks after). It synthesizes memories from your history in the background — so some entries now appear without you asking and without a Memory updated badge. That is the single biggest reason today’s “it remembered something I never told it” and “it didn’t show me it saved anything” confusion exists.

When Memory “doesn’t work,” you’ve almost always hit one of the cases below, in rough order of frequency.

Which bucket are you in?

Symptom	Most likely cause	Jump to
It never remembers anything, any chat	A memory toggle is off	Cause 1
Forgot a specific session’s facts	You were in Temporary Chat	Cause 2
Remembers some things, not others	The statement never triggered a write	Cause 3
Stopped saving new things recently	Memory store is full (~1,200–1,400 words)	Cause 4
Remembers on web, not in a GPT/Project	Custom GPT / Project has separate memory	Cause 5
Worked last week, not today	Workspace switch or sync lag	Causes 6–7

Common causes

1. A memory toggle is off (and there are now two)

Settings → Personalization → Memory exposes two switches as of June 2026: Reference saved memories and Reference chat history. Both are on by default, but privacy-minded users (or an org policy) may have turned one off.

Important gotcha: turning off Reference saved memories also disables Reference chat history. So if the top switch is off, nothing personalizes at all, even though the second toggle may still look enabled.

How to verify: Settings → Personalization → Memory and confirm both toggles are ON. Also confirm the current chat is native ChatGPT, not a Custom GPT (its name shows in the top-left when you’re inside one).

2. You’re in a Temporary Chat

Temporary Chat by design does not read or write either memory layer, and it doesn’t save the conversation to history. It has a dashed border and a Temporary chat label at the top. Saying “remember this” inside one gets a polite yes from the model, but nothing is stored.

How to verify: look for the Temporary chat label at the top of the conversation. If it’s there, start a normal chat instead.

3. What you said never triggered a write

Not every sentence is saved. The model (and now the background synthesis pass) decides “is this a durable fact / preference / long-term detail?” before writing. Vague tone, transient context, one-off questions, and jokes don’t 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: for an explicit save, start with “Remember that…” When a save actually happens, the reply shows a Memory updated line near the bottom. No line = nothing was written that you can see. Note: with Dreaming V3, background synthesis can add entries silently, so the absence of the badge no longer guarantees nothing was saved — confirm in Manage memories.

4. The memory store is full

Saved memories have a soft cap. As of June 2026 it’s best described as a total budget of roughly 1,200–1,400 words across all entries (not a fixed entry count), and OpenAI tunes it. Once you hit the budget, ChatGPT stops adding new saved memories until you delete some, and may surface a “memory full” notice.

How to verify: Settings → Personalization → Memory → Manage memories and skim the list length. If it’s long and dense, you’re likely at the budget.

5. The Custom GPT or Project has its own memory

Two separate scopes trip people up:

Custom GPTs have their own distinct memory and do not read your main-account saved memories. A memory-enabled GPT only works if you also have memory on for your account, and your saved memories are never shared with the GPT’s builder.
Projects keep a separate Project Memory that persists within that Project only. Main-chat memories don’t bleed into a Project, and Project memory doesn’t bleed back out.

How to verify: check the top-left for a GPT name, and check whether you’re inside a Project (the Project name shows in the sidebar/header). If so, the relevant memory lives in that scope, not your main list.

6. Memory was written and read in different workspaces

On Team / Enterprise accounts, the Personal and Company workspaces have separate memory stores. Telling it “remember X” in Personal, then switching to Company, means it won’t know X.

How to verify: the dropdown next to your avatar shows the active workspace. Both conversations must be in the same one.

7. Cross-device or cross-client sync delay

Memory can take a while to propagate between web and the apps. Something you just wrote on web may not appear in the app’s memory list yet.

How to verify: write something on web, then immediately open the app’s memory list — if it’s missing, it likely hasn’t synced yet.

Shortest path to fix

Run these in order; the first three take about 30 seconds each.

Step 1: Confirm both toggles and that you’re not in Temporary Chat

Settings → Personalization → Memory → both Reference saved memories and Reference chat history ON.
Confirm the current chat has no Temporary chat label at the top.
Confirm you’re in native ChatGPT, not a Custom GPT or Project.

Step 2: Open the memory list and read it

Settings → Personalization → Memory → Manage memories. This is the canonical list of explicitly saved facts. Check three things:

Is the fact there? No entry = it was never saved as a saved memory.
Conflicts? Two contradictory entries on one topic (“prefers short” + “prefers detailed”) and the model effectively picks one. Delete the wrong one.
Stale? Old preferences that no longer apply. Delete them.

One caveat as of June 2026: under Dreaming V3 this list is a summary and “does not necessarily include everything ChatGPT may remember,” because Reference chat history has no visible list. If the saved-memories list looks right but behavior is still off, the implicit layer may be carrying old context — see Step 5.

Step 3: Force an explicit write and watch for the badge

Remember that I prefer responses under 200 words unless I ask for detail.

After sending, look for the Memory updated line near the bottom of the reply. Present = saved. Missing = the model decided not to save it; reword it as a durable fact (“I am…”, “I always…”, “My X is…”) rather than a request.

Step 4: Free up space if you’re at the budget

If the list is long and new saves aren’t sticking:

Bulk-delete entries you no longer need (click the trash/X on each, or Clear all to wipe everything).
Re-state your core profile in one message (“I am X, I use Y, I want Z”) so the important facts get re-saved cleanly.

Step 5: Reset the implicit layer if behavior won’t change

If ChatGPT keeps acting on old context that isn’t in Manage memories, that’s Reference chat history. To clear it:

Turn Reference chat history OFF, then ON, in Settings → Personalization → Memory to stop it drawing on old chats.
For a clean slate, delete the underlying conversations, then turn the toggle back on. Note OpenAI may retain logs of deleted memories for up to 30 days for safety and debugging.

Step 6: Pin permanent things to Custom Instructions

For anything that absolutely must persist, use Settings → Personalization → Custom Instructions (the “What should ChatGPT know about you?” and “How should ChatGPT respond?” fields). These are baked into every new conversation’s system prompt and don’t depend on a memory write firing — more reliable than memory for core facts.

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 languages

Step 7: Match the workspace and wait out sync

Team users: verify the write and the read happen in the same workspace (avatar dropdown).
App not synced: force-quit the app, relaunch, and wait 1–2 minutes before re-checking the memory list.

How to confirm it’s fixed

Save a unique test fact: Remember that my test codeword is PURPLE-OTTER.
Confirm the Memory updated line appears, and that the entry shows in Manage memories.
Start a brand-new normal chat (not Temporary, not a GPT/Project) and ask: What's my test codeword?
If it answers PURPLE-OTTER, memory is writing and reading correctly. Delete the test entry afterward.

Often misdiagnosed as

“It remembered me last week but not today” usually means you slipped into a Temporary Chat, a cross-device hand-off hadn’t synced, your workspace switched, or you were inside a Custom GPT / Project with its own memory scope — not that memory “broke.”

Prevention

Treat saved memories like a small notes file — review Manage memories monthly and remove stale or conflicting entries.
Don’t expect cross-account or cross-workspace memory transfer — one account/workspace is one memory pool.
Pin truly important facts to Custom Instructions; it’s more reliable and doesn’t consume your memory budget.
Keep project- or client-specific context in Projects, not main memory, so it doesn’t leak across unrelated chats.
Check the top of a new chat before doing work you want kept — confirm there’s no Temporary chat label.

FAQ

Why does ChatGPT remember things I never told it to remember? As of June 2026, the Dreaming V3 update synthesizes memories from your past conversations in the background, without you saying “remember this” and without a Memory updated badge. Review and prune these under Manage memories.

There’s no “Memory updated” badge — does that mean nothing saved? Not necessarily anymore. The badge fires for explicit saves, but background synthesis can add entries silently. Always confirm in Settings → Personalization → Memory → Manage memories.

What’s the difference between “Reference saved memories” and “Reference chat history”? Saved memories is the explicit, editable list you can see and delete. Reference chat history is implicit recall from past chats with no visible list. Turning off saved memories also turns off chat history.

How much can ChatGPT remember? The saved-memories store is a soft budget of roughly 1,200–1,400 words total as of June 2026 (OpenAI tunes it). When it’s full, new saves stop until you delete some.

Does memory work in Temporary Chat, Custom GPTs, or Projects? Temporary Chat uses neither memory layer. Custom GPTs have their own separate memory and don’t read your main list. Projects have their own Project Memory. None of these read or write your main account’s saved memories.

Why doesn’t my new fact show up in the app right after I saved it on web? Memory can take time to sync between web and the apps. Force-quit and relaunch the app, wait 1–2 minutes, then recheck Manage memories.

Tags: #ChatGPT #Debug #Troubleshooting