Analysis · 2026-10-13 · AI Chat Rank Editorial

Why AI Companions Forget: Context Windows, Memory Slots and Long-Term Recall Explained

The most common complaint about AI companion apps, across nearly every product on our leaderboard, is some version of "she forgot what I told her". It's not a bug in one app. It's the way large language models work, and once you understand the mechanics you can choose an app and a tier that match how you chat, and work around the rest.

The model has no memory of its own

A language model doesn't store your conversation between replies. Each time you send a message, the app assembles a prompt and sends it to the model. The prompt contains the character's description, any persona you've set, any saved memories, and as much recent conversation as fits. The model reads all of that, writes one reply, and keeps nothing. The next reply starts from a freshly assembled prompt.

So "memory" in a companion app is really a question of what the app chooses to put back into the prompt every time. That gives you two levers: how much text fits (the context window) and how cleverly the app decides what to include (summaries and memory features).

Context windows, in plain numbers

The context window is measured in tokens, which are chunks of text of roughly three to four characters each in English. SpicyChat's documentation gives one of the clearest public worked examples. On its free tier, the window is 4,096 tokens. After the character definition and instructions take their share, there's room for about 3,000 tokens of conversation. At about 150 tokens per message, that's roughly twenty recent messages. Message 21 back simply isn't in the prompt any more, so the model can't know it.

Published window sizes from apps on our board:

Two caveats. First, the window you get is the smaller of your tier's limit and the model's own limit. SpicyChat notes that on a top-tier plan, picking a model capped at 8K gives you 8K. Second, long, descriptive replies fill the window faster. A slow-burn story with 300-token replies fits fewer turns than quick back-and-forth chat.

Long-term memory: summaries, pins and slots

Because no window is infinite, apps add a second layer that carries selected facts forward after the original messages have dropped out. The approaches differ:

• Automatic semantic memory. The app extracts details it judges important and re-inserts relevant ones when they seem to matter. SpicyChat's Semantic Memory 2.0, available to subscribers from True Supporter up, works like this.
• Core memory slots. These are a fixed number of facts you or the AI can pin. Erogen describes them as long-term memory, with 8 slots on Silver and 24 on Platinum. The free tier has none, which is why Erogen's large free message allowance still forgets.
• Editable memory panels. Nectar AI lets you write custom memories of up to 200 characters and star up to ten as high importance. Candy AI's Premium Plus lists 20 memory slots. Swipey AI says you can see or delete every memory it stores.
• Layered systems. Nomi AI is the app independent long-term testers credit most consistently with recall weeks or months later. They describe short, medium and long-term layers that resurface details naturally rather than reciting them.

The trade-off is control against effort. Automatic systems are effortless but opaque. When they get something wrong you often can't see why. Manual slots are reliable for what you pin and blind to everything else.

Thread memory versus character memory

A subtle point that catches many users is where memory is stored. In several apps it belongs to the conversation, not to the character. A 2026 hands-on review of Joi AI found that planted facts came back on day fourteen inside the running thread, but not in a fresh conversation with the same character. Nectar AI's docs say memories stay with the conversation where you saved them, and that Reset deletes them. If you like starting new scenes, look for an app whose memory follows the character, or keep one long thread per relationship.

Browser-based apps without accounts add another dimension. Desire AI, our featured partner, says a returning visitor in the same browser can have their saved conversation restored, but that memory is not guaranteed across devices or after clearing browser data. That's the honest cost of not requiring an account.

Why memory often feels worse than the numbers suggest

• Long character cards eat the window. A character with a 1,500-token backstory and long example dialogues leaves less room for your conversation. SpicyChat's docs advise keeping definitions concise for exactly this reason.
• Summaries lose texture. A memory that says "user has a dog" doesn't preserve the joke you made about the dog refusing to walk in the rain.
• Model switching resets style. Apps with many models, such as Nectar's six or CrushOn's tiers, let you change model mid-chat. The new model reads the same prompt but may weigh it differently.
• Conflicts build up. Even strong systems can mix up details over months. One long-term Nomi tester reported two stories about a sibling being blended around the three-month mark.

Practical habits that make any app remember better

1. Open with a two-line recap when you return after a break. It costs a few tokens and puts the key facts back in the window.
2. Pin the facts that matter most if your app has slots or a memory panel. These include names, relationship status and a recent turning point in the story.
3. Keep one long thread per relationship in apps where memory is stored per conversation.
4. Trim the character card. Describe behaviour in a few sentences rather than writing a novel.
5. Restate names when a scene changes. Models drop minor names first.
6. Pick the tier for your style. Quick chatters can live with 4K to 8K. Long roleplayers should budget for 16K or more, or for real long-term memory.

Memory and privacy are two sides of one feature

It's worth remembering that every remembered detail is stored data. The better an app recalls your job, your friends' names and your bad days, the more personal information it holds on its servers. Several apps on our board say they use conversations to improve their models. Joi AI's privacy policy explicitly lists facts about your life, interests and the people you mention as data collected to personalise chat. Good memory and good privacy aren't incompatible, but they need a deliberate choice. Check whether you can view and delete stored memories, as Swipey AI and Nectar AI allow. Find out whether deleting your account removes them. And keep identifying details, such as surnames, employers and addresses, out of anything you'd want remembered. A companion can know your dog's name without knowing your street.

Which apps to look at if memory is your priority

On our scoring, memory carries 25% of the weight. That's why Nomi AI leads the independent apps at 4.2. Erogen AI and Nectar AI offer the most transparent, user-controllable long-term memory, but only on paid tiers. SpicyChat AI publishes its numbers more clearly than anyone, which makes it easy to see exactly what you're buying. Apps whose recall is weak or undocumented, or locked behind their most expensive plan, score lower on our leaderboard for that reason.

Memory is improving quickly across the category, and prices change monthly. Check each vendor's current plan page before you choose a tier. Our figures were checked in October 2026.