What 'Memory' Means in AI Companion Apps

“It remembers you” is one of the most common promises in AI chat and companion app marketing. It is also one of the vaguest. The word “memory” can describe at least three different mechanisms, and they behave very differently when it comes to forgetting, editing and privacy.

This guide separates them, using the official documentation that AI companies publish about their own products. It does not rely on hands-on use of any app, and features change often, so check each product’s current help pages before relying on any detail here.

The three kinds of “memory”

Most products combine all three. Knowing which one is at work explains a lot of confusing behaviour.

The context window: short-term “working memory”

Language models do not remember in the human sense. Each reply is generated from the text the app sends to the model at that moment. Anthropic’s documentation describes the context window as “all the text a language model can reference when generating a response, including the response itself”, and calls it a “working memory” for the model, distinct from the data the model was trained on.

Two points from that documentation matter for companion apps:

• Everything competes for space. System instructions, the character description, every earlier message and the reply itself all count toward the window.
• More is not automatically better. Anthropic notes that as the token count grows, accuracy and recall can degrade.

When a long conversation no longer fits, apps have to drop or compress older parts. That is the most common reason an AI “forgets” something you said an hour ago: it is no longer in the text the model is shown.

Long-term memory: what apps save between chats

To remember across conversations, an app has to store information somewhere and re-insert it later. How that works varies.

ChatGPT as a documented example

OpenAI’s help centre is one of the more detailed public descriptions of a memory system. According to it:

• When Memory is on, ChatGPT can draw on sources such as past chats, saved memories, custom instructions and, depending on plan and region, files and connected apps.
• Memory “does not retain every detail from every conversation”. ChatGPT decides which information is relevant.
• Saved memories are details you ask it to remember, or that it saves as useful context, and they are stored separately from chat history.
• Reference chat history lets it use relevant information from past conversations, and what it derives can change over time.
• A memory summary gives a high-level view, but OpenAI says it does not necessarily include every detail or source Memory can reference.

ChatGPT is a general assistant rather than a companion app, but its documentation shows the pattern many products follow: an explicit list of facts plus a less visible layer built from your history.

Replika as a companion-app example

Replika’s help centre describes its memory as “built on layers”. Some memories are visible in a Memory tab; others belong to “a deeper system that draws on patterns from your overall conversation history”. It also says Replika remembers profile details you provide, such as your name, and usually picks up things like favourite books, music and the people you mention.

The two documented systems share a structure: a visible, editable list, and a background layer you cannot fully inspect.

User-edited memory: the part you control

Both companies document ways to see and change what is stored:

• Ask directly. OpenAI says you can ask ChatGPT what it remembers about you. Replika suggests asking “What do you know about me?”
• Review a list. Replika’s Memory tab lets you scroll through saved memories, tap one to read it fully or remove it. ChatGPT offers a memory summary or saved-memory list, depending on the account.
• Add or correct. Replika lets you create memories manually, used alongside auto-generated ones. OpenAI lets you correct the memory summary or tell ChatGPT something is outdated.
• Suppress. OpenAI’s “Don’t mention this again” reduces future references but, in its words, “does not delete the original source”.

Replika also notes that deleting memories “may affect your Replika’s ability to recall certain details”, so there is a trade-off between control and continuity.

Why your AI companion forgets

Putting the mechanisms together, forgetting usually has one of these causes:

1. The conversation outgrew the context window, so older messages were dropped or summarised.
2. The detail was never saved to long-term memory, because the system judged it unimportant.
3. A summary replaced the detail. Summaries lose specifics.
4. Memory is off, or limited by plan. OpenAI says memory features can vary by plan, region, platform and workspace.
5. You deleted or suppressed it, or started a mode that does not use memory.

Worked example

Suppose you mention early in a long chat that you are training for a 10 km run in May. Days later, the companion asks what sport you play. Likely explanations, in order: the original message is no longer in the context window, and the background memory saved something vaguer, such as “likes fitness”. The fix is to check the memory list, and if the detail is missing, add it yourself in a short, specific form.

Privacy implications

Memory is useful precisely because it stores personal information. That has consequences worth understanding before you share anything sensitive.

Deleting a chat may not delete the memory. OpenAI states that “deleting a chat alone does not necessarily delete a separate saved memory created from that chat”. To remove something fully, it says to delete the saved memory and the chat where you first shared it, plus any other sources such as files or connected apps.

Deletion is not always instant. OpenAI says deletion and memory updates can take time to propagate, and that it may retain logs of deleted saved memories for up to 30 days for safety and debugging. If you turn off Reference chat history, information remembered from past chats is scheduled for deletion within 30 days.

Turning memory off is not the same as deleting history. OpenAI notes that turning off Memory does not delete past chats, and if you turn it back on, new memories may be created from chats still in your history.

Private modes have rules. OpenAI says temporary chats do not create or update memories, but saving a temporary chat converts it to a regular one.

The privacy policy is the full list. Replika’s help centre points users to its Privacy Policy for a complete list of the information it retains. On Android, Google Play’s Data safety section is where developers describe what data their app collects and shares, though Google notes practices can vary by use, region and age.

Memory privacy checklist

• I know whether the app has long-term memory and whether it is on by default.
• I have found where to view and delete stored memories.
• I know whether deleting a chat also deletes memories created from it.
• I know whether there is a temporary or private mode, and what it does.
• I have read the privacy policy section on retention and deletion.
• I have not shared passwords, financial details or health information I would not want stored.

For a fuller list, see our AI companion privacy checklist.

How to judge memory claims in reviews and ads

When a product claims “unlimited memory” or “never forgets”, ask which kind it means. A larger context window, a saved-facts list and a background summary are not equivalent. Our guide on how to evaluate an AI companion app includes memory as one of several criteria. Memory features are also a common paid upgrade, so check whether the free tier includes them: see free vs paid AI companions and AI companion subscription costs. If you are comparing a specific product, our Candy AI review and Candy AI alternatives pages note what each documents.

Bottom line

“Memory” in AI apps is usually three things stacked together: a context window that holds the current conversation, a long-term store the app fills in automatically, and a list you can view and edit. Forgetting is normal when conversations outgrow the window or details never reach the long-term store. Deleting a chat may not delete what was learned from it, and deletion can take time. Before sharing anything personal, find the memory controls, read the retention section of the privacy policy, and use a temporary mode where one exists.

For how we source and label guides like this one, see how AITried reviews AI products. If you are paying for memory features, the subscription calculator shows the yearly cost.