Plain-English explanation
Character memory is the product experience described as "my AI remembers me." It is not a specific technical method: an app might implement it by keeping enough conversation history inside the model's context window, by maintaining a separate stored summary of key facts, or by using a full RAG system that retrieves relevant past details on demand. From the user's side, all of these can look similar: the character brings up something you mentioned before, or reacts consistently with established history.
How it works
Character memory implementations generally fall into a few tiers, roughly in order of sophistication:
- In-context only. The app keeps as much recent conversation as fits in the context window. This is simple but limited: once the conversation grows past the window, older detail is lost unless captured some other way.
- Summarization. The app periodically compresses older conversation into a shorter summary, which is then kept in context going forward, trading detail for space.
- Structured fact storage. The app extracts specific facts (your name, preferences, key events) into a separate profile that's re-inserted into every conversation regardless of how long ago it came up.
- Retrieval-based (RAG). The app stores past conversation in a searchable form and pulls back only what's relevant to the current message: see RAG and long-term memory.
Many companion apps combine several of these tiers rather than relying on just one.
Adult-AI use
Character memory quality is one of the most-marketed and most user-visible differentiators between companion apps: it's often advertised directly ("remembers everything about you") and is a major factor in whether an ongoing relationship with a character feels believable over weeks or months rather than resetting every session. It's also a genuine technical challenge: good long-running memory is expensive to build and run well, which is part of why memory quality varies so much between free and paid tiers, and between smaller apps and better-resourced platforms. Platforms such as Nomi and Kindroid market multi-tier memory architectures directly as a product differentiator, with their database profiles documenting the exact tier structures.
How it differs from context window, long-term memory, and RAG
Character memory is the umbrella, user-facing term; the other three are the underlying mechanisms. Context window is the raw technical limit on what a single request can hold. Long-term memory specifically means memory that persists beyond the current context window and session. RAG is one specific technique for implementing that persistence. A companion app's "character memory" feature is typically built from some combination of the three, not a separate fourth technology.
Limitations and misconceptions
- Memory can be wrong, not just missing. A character can confidently misstate or invent a detail about your history together (a hallucination) rather than forgetting it, which can be more jarring than an honest "I don't remember."
- "Remembers everything" is rarely literal. Most implementations selectively retain or summarize rather than storing a perfect, complete transcript indefinitely, both for cost reasons and because feeding a model too much irrelevant history can hurt response quality.