Who Decides When Your AI Assistant's Memory Goes Stale

A teardown of how ChatGPT, Claude, Gemini, Copilot, Muse and Bee decide what to keep about you, what that memory is allowed to influence, how it ages, and who gets to retire it.

Read this first. I went in expecting assistant memory to be a storage problem with a delete button. The companies' own documentation describes something messier. Here are three findings, ranked by how much they should change the way you evaluate these products:

1. Only Claude and Gemini document a clean way to keep memory without using it, and only in bulk. Claude's Pause holds every memory and stops using them. Gemini lets you switch off each input type, or all personal memory for one chat. ChatGPT's and Copilot's chat-history toggles look similar, but they schedule what was learned for deletion within 30 days. Muse and Bee document nothing comparable. For a single memory, ChatGPT is the only product with real controls: you can hide it or lower its priority. Elsewhere, you make a request. None of these controls has an end date. High confidence on what's documented.
2. OpenAI and Anthropic both shipped fixes for outdated memory in 2026, and neither documents when or why a given memory changed. OpenAI rewrites memories in background synthesis. Anthropic updates topics mid-conversation. Neither documents a per-memory date or a stated reason for the change. ChatGPT's version history is the closest thing to an audit trail. High confidence on what's documented. I have not tested whether the rewrites are accurate.
3. No general-purpose assistant shows you, for each memory, where it came from or when it was last confirmed true. Without that link from source to memory, you can't remove everything learned from one chat or one app as a set, which is why deleting a chat or disconnecting an app usually leaves its lessons behind. High confidence on what's exposed. Moderate confidence that the systems store more than they show.

Signal status: developing. The companies named the stale-memory problem this year, but the visible fix, dates and sources on each memory, hasn't shipped.

On sourcing. Everything below comes from help centers, release notes, developer docs and a year of user reports on Reddit and Hacker News, current as of this week. Help pages describe intended behavior, so when I write that a product "does" something, read it as "says it does."

Two traces matter throughout. Provenance runs from source to memory: where did the assistant learn this? It's what would let you delete everything learned from one app. Attribution runs from memory to answer: which memories shaped this reply? It's what would let you check that a suppression held. The table keeps them apart.

Six assistants across four stages of memory

What gets kept, and how far it can travel

The inputs keep widening. ChatGPT builds memory from chats, files and connected apps. Gemini treats past chats, saved details, instructions and connected Google apps as separate inputs, each with its own switch. Muse pulls from conversation, observed preferences, connected apps and a file you can edit by hand.

Most of the controls built around those inputs limit where a memory can travel:

• ChatGPT offers project-only memory, which keeps a project's chats out of your main memory. Locked chats, announced October 20, wall off anything learned inside them, though locking a chat doesn't pull back what ChatGPT learned before you locked it.
• Claude mixes project memory with account memory unless you turn off Use account memory.
• Duet, Anthropic's assistant for working across connected apps, runs on your Claude memory. Its topic channels share that one memory rather than keeping separate ones.

The design effort has gone into scoping memory by place, and that work is real. Scoping by time (was this true until some date, and which date?) barely exists as a user-facing control.

What memory is allowed to steer

Claude's Pause is the cleanest keep-but-don't-use control. It retains your memories while stopping Claude from reading them or writing new ones, across your whole account. Gemini splits the same idea by input: separate switches govern whether past chats and instructions personalize answers, and saved instructions stay until you delete them. Gemini also lets you turn off personal memory for a single chat, which shuts off past chats, instructions and connected apps together; the next new chat defaults back on. Claude's per-chat off switch works only if you choose it before the first message.

Two switches that look like they belong on this list are really deletions with a grace period. Turning off "Reference chat history" in ChatGPT schedules what it learned from your chats for deletion within 30 days. Copilot's chat-history personalization works the same way, and restores the inferred details if you switch back in time. Copilot's saved-memory switch does keep existing memories, but Microsoft describes it as stopping new saves and doesn't say old memories stop steering answers.

For one specific memory, ChatGPT has the only real controls. "Don't mention this again" changes personalization without deleting the source, though OpenAI doesn't promise the memory stops influencing answers altogether. And since October 2025, ChatGPT has ranked saved memories by recency and frequency and let you raise or lower an individual memory's priority. Lowering priority is the closest any product comes to "keep this fact, but lean on it less."

Elsewhere you ask. Microsoft says you can tell Copilot to avoid a specific detail. Google warns that asking Gemini to forget or avoid a topic may not work, and advises deleting the related chats if you need certainty. None of these controls lapses on its own, and I found no product offering anything like "don't use this for six weeks."

Suppression would matter less if you could check that it held. That check is an attribution problem, and no assistant fully shows which memories shaped a given answer. ChatGPT's Sources view lists contributing chats, memories, files or emails, but OpenAI says it may leave some out. Claude cites the past chats it found when it searches, which tells you what it looked up rather than which stored topics shaped the reply. Gemini says it will try to explain.

You've shipped agent systems, where a policy with no execution trace wouldn't survive an enterprise security review. Consumer memory ships one anyway: a rule ("don't use this") with no log showing the rule was followed.

The model holds the clock on aging memories

The products moved most here in 2026, and this is where my starting assumption was most wrong: the companies are not ignoring outdated context.

OpenAI has two mechanisms, eight months apart. The October 2025 ranking pushes memories you haven't discussed lately into the background. The June Dreaming release has ChatGPT rewrite memories in the background as time passes. OpenAI's example: an upcoming trip to Singapore becomes a finished trip, and recommendations shift back to the user's home city. The evaluations showing this works are OpenAI's own.

Anthropic rebuilt Claude's memory in August to update topics during conversations instead of in a daily batch. Its examples are a deadline that moved and a company that renamed itself.

Microsoft updates outdated details in Copilot and shows a "Memory updated" notice that doesn't say what was replaced.

What's missing is a date you can see on each memory. Across the help pages for ChatGPT, Claude, Gemini, Muse and Copilot, I found no consistent display of when a memory was created, last used, last confirmed or due to expire, and no way to set an expiry on a single memory. ChatGPT comes closest. You can sort saved memories chronologically and restore earlier versions, so it keeps dates somewhere, but its documentation doesn't say a creation or last-used date appears on each memory.

The time controls that do exist govern how long records are retained, which is a separate question from whether a fact is still true. Gemini deletes chat records after 18 months by default, with options for 3 months, 36 months or never. Amazon's Bee keeps history by plan, 14 days free or unlimited on Pro, and its documentation doesn't say what happens to extracted facts on day 15.

Bee is the one product where the building blocks show up. Its developer tools expose, per fact, a creation time and a confirmed flag. Both appear only in the developer documentation; I found nothing showing them in the consumer app, and the docs don't explain how "confirmed" changes what Bee relies on. Speculative whether the flag affects behavior at all; the documentation is silent.

Why visible dates matter even when the model does the updating. A trip has a date attached. Your job, diet, health situation and home city don't. For those, the model has to infer from conversation that something stopped being true, and it can miss in either direction, dropping a fact that still holds or holding one that doesn't. When it misses, you end up disputing undated text that has already been rewritten. Moderate confidence: rewriting in place makes memories right more often and harder to audit when they're wrong.

My Figma teardown raised the same risk for specs in a shared file, which get treated as current because they're present. A fact sitting in memory gets the same treatment, whether or not anyone has checked it lately.

User evidence is thin. I searched a year of Reddit, Hacker News and product forums and found one clean case. A Gemini user mentioned an upcoming move, and Gemini kept referring to everything being boxed up after the move had happened. Other complaints, such as an assistant mentioning someone's job too often or old client details surfacing in new client work, are about relevance rather than staleness. Speculative: without dates on memories, users may not recognize a memory as outdated, so they report it as creepy or irrelevant instead.

Deleting the source leaves the lesson behind

On deletion, the help pages are unusually candid. Deleting or letting a Claude chat expire doesn't remove the memories it produced; you delete those separately. Meta warns that deleting something from Muse may leave behind what Muse learned from it. Gemini requires you to delete the relevant chats and disconnect the app, because either one can reintroduce the information.

Duet says it most plainly. From Anthropic's documentation:

So cutting an app's access and removing its influence are separate operations. As a default, that's defensible. The memory is about you, and plenty of people would be annoyed if disconnecting Gmail erased what Duet knew about their week. The problem is that you don't get to choose. Duet's help pages show no list of memories derived from a given app, no filter for them, and no way to remove them as a set. You delete topics one at a time or reset everything.

My Manus piece separated three things an agent product holds for you: connection settings, credentials and task state. Memory is a fourth, and by the documentation's own account it outlasts the other three.

Users see the memory without its source. What the systems store is less clear, and some link evidently exists. ChatGPT's Sources view can name the chats or files behind an answer, and Gemini's two-step removal implies it rebuilds memory from whatever sources remain. Moderate confidence: the source-to-memory link is at least partly stored, but no product exposes it as a field on each memory or uses it to delete a memory along with its source.

Where memory aging already works

The design exists outside consumer assistants:

• GitHub Copilot's memory deletes unused facts after 28 days; the timer can reset when a fact is validated and used. Facts about a repository are verified against the current code. Personal preferences get a best-effort check.
• Generative Agents (Park et al., UIST 2023) stamps every memory with a creation time and a last-accessed time, then retrieves by blending recency, importance and relevance, with memories losing weight the longer they go unretrieved.
• CRM software has carried "last verified" fields for years so a salesperson knows how far to trust a phone number.

Coding tools got there first because they have a current source of truth to check against: the codebase. For a consumer assistant, the source of truth is you, and checking each memory means asking you whether it's still true. That adds friction, which cuts against the effortless memory these products sell. Moderate confidence: consumer assistants hide memory dates and sources as a product choice, since the systems already keep some of both.

Friction may not be the only incentive. Speculative: the documented defaults also make memory a retention asset for the platform. Memory survives app disconnection, carries from Claude into Duet, and can't be retired by source; Gemini's draws on Google's own apps. Each of those choices makes the accumulated picture of you more valuable inside one platform. If that's the logic, expect per-memory sources to arrive slowly, because they're what would let you take that picture apart.

Verdict

Incremental, not structurally new. Dreaming is the most significant move in this space, ahead of Anthropic's August rebuild, because it acts on the passage of time itself. Anthropic's examples (a moved deadline, a renamed company) are changes someone reports; OpenAI's is a trip that ends without anyone saying so. Still, Dreaming brings techniques research already had, recency weighting and background rewriting, to consumer assistants, and keeps its timekeeping out of view.

Ranked by question:

• Suppressing one memory: ChatGPT leads with hide and down-weight. Copilot's ask-to-avoid is second. Nobody offers a time-bounded suppression.
• Visible aging: Bee's developer tools lead, followed by ChatGPT's chronological sort and version history.
• Honest retirement: Claude and Duet document the gap most plainly. Gemini puts the most work on the user to fully remove something.

A structurally new design would be a memory that carries its source and its last-confirmed date and can be retired by either. GitHub Copilot and Bee's developer tools each hold part of that, and no general-purpose assistant exposes it to users. The companies have accepted that facts go stale; so far they've handled it by letting the model decide when that happened, without showing you the decision.

Watch for:

• The first assistant to show a source and a "last confirmed" date on each memory.
• Bee documenting its confirmed flag in the consumer app, not just in developer tools.
• "Don't mention this again" becoming a checkable setting with a record that it held.

For your search. When you talk to teams hiring for memory or personalization design, ask early whether design owns the rules (what ages, what requires confirmation, what gets deleted with its source) or only the settings page. Those rules are the attribution and tracing problem you worked on with enterprise agents at TinyFish, moved into a consumer product. If the answer is the settings page, treat it as a fit warning: the team has already placed the hard part outside design.