Your AI Assistant Keeps Forgetting Things. Here's How to Fix That

Your AI assistant isn’t broken: it’s running three separate memory systems, and each one fails differently. ChatGPT, Claude, and Gemini all work the same basic way: a context window (what the model can read in the current chat), saved memory (facts carried between chats), and project instructions (rules and files pinned to a workspace). When your AI assistant memory fails, one of those three layers dropped the ball. Match the fix to the layer: flip the right toggle, say “remember this,” or move the work into a Project, and the forgetting mostly stops.

Below: how each memory system works today, a side-by-side comparison, an eight-step fix, and the habits that keep any of them on track.

Why does your AI assistant forget things? Three systems, three failure modes

Forgetting mid-conversation is almost always the context window running out. Forgetting between chats is a memory-settings problem. Forgetting inside a Project is usually a scoping problem. Three different systems, three different fixes, which is why switching everything off and on again rarely solves anything.

Here’s the vocabulary, because the three get confused constantly:

• Context window is the text the model can actually read at one time: current messages, replies, attached files. Past that limit, the oldest parts stop shaping responses. You can scroll up and read them; the model effectively can’t.
• Saved memory is the long-term layer: facts the assistant writes to a persistent store and loads into future chats. You usually write to it by saying “remember this,” and it accumulates whether or not you’re watching.
• Project instructions are standing rules and files attached to a workspace, applied to every chat inside it: a briefing the assistant re-reads every time.

The windows are enormous now: GPT-6.1 Sol at 1,050,000 tokens of context with 128,000-token outputs (OpenAI’s API docs); all four current Claude models (Fable 5.1, Opus 5.5, Sonnet 5.5, Haiku 5.5) at 1M with 128K output; Gemini 4 Argon at 1M with 65K output.

A million tokens is roughly 750,000 words (a conversion, not a vendor figure) and every current flagship reads that much in one go. It still isn’t memory. Close the chat, and none of it carries over unless the app wrote it down somewhere.

Two honest caveats: none of the vendors publishes exact per-chat context limits for the consumer apps, so the API specs are the model’s ceiling, not a promise about the app. And in very long chats, early material tends to lose influence gradually rather than vanishing cleanly: the docs don’t quantify that, but it’s the failure pattern users report most.

How ChatGPT memory works right now

ChatGPT (OpenAI) splits memory into two switches (“Reference saved memories” and “Reference chat history”) plus a newer Memory summary view, all under Settings → Personalization → Memory. Exact controls vary by account and plan, per OpenAI’s help center.

The two layers behave differently:

• Saved memories are details you explicitly ask ChatGPT to remember, or it saves on its own as useful context, stored separately from chat history.
• Reference chat history lets ChatGPT pull relevant bits from past conversations on its own; what it derives “can change as ChatGPT updates what is most useful to remember.” Turning it off schedules that derived info for deletion within 30 days, and turning off saved-memory reference also switches off chat history.

The Memory summary (a high-level, correctable view of what ChatGPT thinks it knows) is rolling out gradually; some accounts see it, some the older Saved memories list, some both.

Quirks worth knowing before you trust it:

• Deleting a chat does not necessarily delete a saved memory created from that chat. OpenAI may also retain logs of deleted memories for up to 30 days for safety and debugging.
• Temporary chats create and update no memories, even with personalization on. Save a temporary chat, though, and it converts to a regular one.
• There’s no stated numeric cap on how much chat history the memory system can draw on.

Custom instructions are a third channel: standing rules applied to new conversations on every plan, capped at 1,500 characters on Free and Go and 5,000 on paid plans. Pricing runs Free, Go $8/month, Plus $20/month, Pro $100-$500/month (November 2026).

Projects are the biggest lever most people ignore. A Project groups chats, files, and its own instructions; file caps are 5 on Free, 25 on Go and Plus, 40 on Pro, Business, Enterprise, and Edu, with no limit on the number of projects. Project-only memory limits recall to conversations inside the project: set per project under ••• → Project settings → Memory, since there’s no global switch. Project instructions override your global custom instructions inside that project, and shared projects are automatically project-only, blind to members’ personal memories.

How Claude memory works right now

Claude (Anthropic) remembers by saving individual topics as you chat, rather than summarizing after the fact. Memory is on by default for Free, Pro, Max, and Ultra on web, desktop, and mobile; Team and Enterprise admins control it org-wide.

Nothing is hidden in Claude’s implementation. Settings > Your Claude > Manage shows a list of Topics (readable, editable, deletable) and saying “remember this” saves a topic directly.

Claude is also the most explicit about what it won’t remember:

• Incognito chats (the ghost icon) are never saved to history or memory, on every plan.
• Sensitive topics (health, race, religion, politics, and similar) are excluded from memory by default, with an opt-in if you want them included.
• Some categories are never saved even if you ask: government ID numbers, criminal history, financial account numbers, and immigration status.

The off switch offers two choices that people mix up: Pause memory keeps everything stored but stops using or adding to it, while Reset memory permanently deletes all memories, including project memories, with no undo.

Each Project keeps its own memory, separate from your account-level memory, and you can disable “Use account memory” on any project page for full isolation. Chat search on paid plans (toggle “Search and reference chats”) lets Claude search non-project chats and cite them with links back to the originals. For recurring work, the shape is: project instructions + knowledge-base files + per-project memory, all visible. Free users can create up to five projects; paid plans expand knowledge capacity via retrieval, and a newer beta (rolling out to select Pro, Max, and Ultra users who use Claude Code) treats a whole project as one long-running conversation.

How Gemini memory works right now

Gemini (Google) ships two things both called “memory”: Instructions for Gemini (standing rules anyone can add) and automatic Memory of past chats, which is US-only and gated behind a personal Google Account, age 18+, Keep Activity on, and an active Google AI plan; instructions work for everyone. Both live under Settings → Personal Intelligence.

The instructions layer is simple: type “Remember that I am vegetarian,” and Gemini stores it. Manage the list at gemini.google.com/saved-info or in settings.

The automatic layer is where it gets messier:

• The Memory summary view is currently limited to US users with an active Google AI plan, in English, on the mobile and web apps. Ask “What do you know about me?” to see it, and check Memory history to see when entries were added or changed.
• Deletion runs through Gemini Apps Activity: delete the chats containing the info, and note Google’s own warning that there’s a short delay before Gemini stops using deleted material.
• Connected apps need double removal: delete the chats and disconnect the app. Doing only one leaves the info reachable through the other.
• Memory doesn’t apply in Gems or Live chats. Google is also transitioning Gems toward “skills,” per a September 2026 Workspace announcement.

Temporary chats (added August 2025) don’t appear in history and aren’t used for personalization or training. Google’s tiers run AI Plus $4.99/month, AI Pro $19.99/month, AI Ultra from $99.99/month; automatic memory needs a paid plan plus US availability.

ChatGPT vs Claude vs Gemini: how AI assistant memory compares

All three now do the core trick (carrying context between chats) but they differ sharply on who gets the feature, how much you can see, and how neatly it scopes to projects.

Capability	ChatGPT	Claude	Gemini
Automatic cross-chat memory	Yes, via “Reference chat history” toggle	Yes (topic-based, on by default on Free/Pro/Max/Ultra	Yes) US-only; needs personal account, 18+, Keep Activity, and a Google AI plan
Explicit “remember this”	Yes (saved memories	Yes) saves a topic on request	Yes, saved info/instructions (all users)
View and edit what’s stored	Memory summary + Manage list	Settings > Your Claude > Manage topics	Memory summary + saved-info page
Per-project scoping	Project-only memory, set per project	Each project keeps its own memory	Gems/skills instructions; no per-project memory
Private session mode	Temporary chats create no memories	Incognito chats never saved	Temporary chats not saved or used for training
Memory on the free plan	Yes	Yes, on by default	Instructions yes; automatic memory needs a paid plan (US)

Across the three help centers, ChatGPT is the most automated: it builds a profile with the least input, convenient and slightly unnerving. Claude is the most inspectable and most conservative, with clear categories it refuses to store. Gemini is the most fragmented: instructions, automatic memory, and a summary live in three places, but its controls tie into Google’s account-wide activity settings. None can read another’s memory; every silo is self-contained.

The fix: eight steps to repair your AI assistant’s memory

Work through these in order. Most people find the problem in steps 1-4.

1. Name the failure. Forgot something mid-chat? Context window. Forgot between brand-new chats? Memory settings. Forgot inside a Project? Scoping: check whether the project is set to project-only memory.
2. Check the toggles. ChatGPT: Settings → Personalization → Memory, turn on Reference saved memories and Reference chat history. Claude: Settings > Your Claude > “Generate memory from chats.” Gemini: Settings → Personal Intelligence → Memory (plus Keep Activity on, personal account, and a Google AI plan where required).
3. Say “remember this” for anything you want kept. Explicit saves are the one thing you directly control on all three. State facts with context: “Remember: our client Freightline uses British English spelling in all deliverables.”
4. Audit what it already knows. Open ChatGPT’s Memory summary or Saved memories, Claude’s Manage topics list, or ask Gemini “What do you know about me?” Delete anything wrong or stale: this catches the bad memories polluting new chats.
5. Move recurring work into a Project. Put the standing rules in project instructions, attach the key files, set memory scoping deliberately. It beats general memory for anything with a client, codebase, or house style.
6. Keep a briefing document for high-stakes work. One file with your goals, constraints, voice notes, and current status; paste it at the start of important chats and fold in a session-end summary. No vendor documents this pattern, but heavy users swear by it, and it’s the most reliable habit here.
7. Re-feed context when a chat runs long. Re-paste the key constraints, or start fresh with a short summary of where things stand.
8. Delete wrong memories at the source. Remove the memory entry, the originating chat where relevant, and any connected app (Gemini’s double-removal rule), then expect a propagation delay: both OpenAI and Google warn changes take time to apply everywhere.

Habits that keep any assistant on track

The settings only get you halfway; the rest is routine. Five habits that hold up across all three assistants:

• Start every recurring engagement with a Project, not a chat: instructions and files beat remembered fragments.
• End long sessions by asking for a handoff summary you can paste into the next chat.
• Audit your memories monthly; wrong facts you never notice are worse than forgotten ones.
• Use temporary or incognito chats for anything sensitive: tax documents, health questions, client secrets.
• Store each fact in one place. The same rule living in custom instructions, saved memory, and a project file creates conflicts that look like “forgetting.”

What your assistant remembers: and what you can do about it

Every assistant now gives you real controls, but they’re scattered, and the details matter. Claude excludes sensitive topics by default, never stores a defined set of categories, and lets org owners turn memory off for everyone: permanently deleting all memory data on Team and Enterprise. OpenAI keeps Business, Enterprise, and Education content out of training by default; personal accounts depend on “Improve the model for everyone.” Gemini’s temporary chats aren’t used for personalization or training, and everything automatic memory learns ties back to Gemini Apps Activity, which you can delete wholesale.

One caution: consumer AI chats aren’t compliance infrastructure. Handling regulated data: client health records, financial details, anything under legal hold? Talk to a professional before putting it through any of these tools, memory off or not.

What to do next

If you do nothing else: open each assistant’s memory settings today, turn the toggles on, and read what’s already stored. That alone fixes the “it forgot my job/preference/agency name” problem for most people.

From there, match the setup to how you work. Casual users need nothing beyond toggles and an occasional audit. Businesses running client work should spend the hour on Projects (instructions, files, scoped memory) because it pays back within a week. Privacy-minded users should lean on Claude’s visible topics or Gemini’s temporary chats, and treat ChatGPT’s profile-building as powerful but worth reviewing regularly. Whichever you pick, keep the briefing document. The assistants will keep changing; the file that re-explains you never goes out of date.

FAQ

Why does my AI assistant keep forgetting what I told it?

Most forgetting comes from the context window: the fixed amount of text the model can read at once. In long chats, early messages lose influence or drop out entirely even though you can still see them on screen. Forgetting between separate chats is different: that’s a memory-settings problem. Match the fix to the failure before changing anything.

What’s the difference between a context window and saved memory?

A context window is short-term: everything the model reads during the current chat, gone when the chat ends. Saved memory is long-term: facts the assistant writes to a persistent store and loads into future chats. Project instructions are a third layer: standing rules and files attached to a workspace and applied to every chat inside it.

Does ChatGPT memory work across all chats?

Yes, when Reference saved memories and Reference chat history are on under Settings → Personalization → Memory. Exceptions: temporary chats create no memories, and projects set to project-only memory can’t reference anything outside that project. Shared projects are always project-only and can’t see members’ personal memories.

How do I see and delete what ChatGPT remembers about me?

Go to Settings → Personalization → Memory. Depending on your account you’ll see a Memory summary, a Saved memories list, or both. Delete individual entries, or use “Delete and turn off memory” to wipe everything at once. Deleting a chat doesn’t necessarily delete a memory created from it, so delete the memory too.

Does Claude remember previous conversations?

Yes. Claude’s memory is on by default for Free, Pro, Max, and Ultra plans and saves individual topics as you chat. View, edit, or delete each topic under Settings > Your Claude > Manage. Incognito chats are never saved, and sensitive topics are excluded from memory unless you opt in.

Does Gemini remember previous chats?

Yes, if you’re 18 or over, sign in with a personal Google Account, have Keep Activity on, and live where the feature has rolled out: automatic memory is US-only and requires an active Google AI plan. Anyone can add standing instructions under Settings → Personal Intelligence → Instructions for Gemini.

Do I need a paid plan to get AI memory features?

Not for ChatGPT or Claude: ChatGPT’s Free plan includes memory and Claude’s is on by default for Free accounts. Gemini’s automatic memory requires a Google AI plan (AI Plus from $4.99/month, AI Pro $19.99) and is US-only. Paid plans generally add longer context, more project files, and chat search.

How do I stop my AI assistant from forgetting important instructions?

Don’t rely on chat memory for anything that matters. Put standing rules in custom instructions (ChatGPT and Gemini) or project instructions, attach key files to a Project, and keep a briefing document you paste at the start of high-stakes chats. Ask for a summary at the end of each session to add to it.

What should I do first if memory still isn’t working?

Check the toggles first: Settings → Personalization → Memory in ChatGPT, Settings > Your Claude in Claude, Settings → Personal Intelligence in Gemini. Then audit what’s stored, delete wrong memories, and remember that deletions take time to propagate. If nothing helps, start a fresh chat and re-supply the context manually.

Sources

• Memory in ChatGPT: OpenAI Help Center
• ChatGPT Custom Instructions: OpenAI Help Center
• Projects in ChatGPT: OpenAI Help Center
• ChatGPT pricing: OpenAI
• GPT-6 announcement and updates (GPT-6 Sol, Luna, GPT-6.1 Sol): OpenAI
• OpenAI API models overview
• Use Claude’s chat search and memory to build on previous context: Anthropic Support
• What are projects? Anthropic Support
• Claude model overview. Anthropic documentation
• Claude release notes. Anthropic Support
• Get personalization with memory of your past Gemini chats: Google Help
• View and manage your memory summary in Gemini Apps: Google Help
• Manage what Gemini remembers about you (saved info). Google Help
• Temporary chats and privacy controls in Gemini: Google blog
• About the transition from Gems to skills: Google Workspace Knowledge
• Gemini models and token limits. Google AI Studio documentation
• Google AI plans. Google One