How AI Companion Memory Works
AI companions combine two distinct memory mechanisms: the context window (everything visible to the model in a single conversation) and long-term memory (a separate database of stored facts retrieved at the start of each new session). Understanding the difference is key to evaluating any companion app.
When you chat with an AI companion, the model only directly "sees" the text inside its context window — a finite chunk of conversation measured in tokens (roughly 0.75 words each). A model with a 128,000-token context window can hold roughly a novel's worth of text in one session, but once you close the app and return the next day, that window is gone. This is why early chatbots felt so amnesiac.
Long-term memory solves this by storing important facts — your name, your job, significant events you've shared — in a separate database. When you start a new session, the app retrieves relevant facts from that store and injects them into the context window as a "memory summary." The model then responds as if it remembers you, because it has just been told the key facts. The quality of this system — what gets stored, how accurately it's retrieved, how much context it includes — varies enormously between apps and is one of the primary differentiators Companion Index measures.
There is an important implication for privacy: anything the app stores in your long-term memory profile may be used to train future versions of the model, sold to third parties, or exposed if the platform suffers a data breach. Apps take very different approaches here: some store fine-grained, user-editable relationship context, while others keep memory shallow and scoped per character. Companion Index measures each app's memory depth as a scored, sourced signal. Apps that allow you to view, edit, and delete your memory profile — and that offer a genuine data-export option — give you far more control over your data. Companion Index tracks data export and account deletion capabilities for every app it reviews.
FAQ
- What is a context window?
- The context window is the maximum amount of text a language model can process in a single conversation turn. Everything outside the context window is invisible to the model unless it has been stored in long-term memory and re-injected. Modern companion apps use context windows of 32,000 to 200,000 tokens.
- Is long-term memory the same as the AI actually learning from me?
- Not necessarily. Most apps implement long-term memory as a retrieval system — they store facts and inject them at session start — rather than as genuine model fine-tuning. True personalisation via fine-tuning is rarer and more expensive. The distinction matters: retrieval memory can be edited or deleted by the user; fine-tuned model weights generally cannot.
- Can I delete what an AI companion has memorised about me?
- It depends on the app. Some platforms offer full memory editing and account deletion with data erasure; others retain data even after account closure. Companion Index tracks this as an observable signal for every reviewed app.