Conversational Memory
Conversational memory is the component of an AI companion's memory system that stores the content, context, and emotional tone of past conversations so that future exchanges can reference and build on them. It is distinct from a session's context window — which resets — and enables the continuity that makes AI relationships possible.
Conversational memory vs a context window
Every AI language model has a context window — the amount of text it can process in a single inference call. Within a session, this window holds the conversation so far. When the session ends, the context window is discarded. A model with a longer context window can hold more of the current conversation, but once you close the tab, nothing is retained.
Conversational memory is different. It is an external storage system — a database — that persists indefinitely across sessions. When you return to a companion days or weeks later, the companion does not start from scratch. It has a record of what you discussed and can retrieve the most relevant pieces when they would help the current conversation. See Long-term Memory for how this retrieval works at scale.
What is stored in conversational memory
Not every word of every conversation is stored verbatim. Storage typically works in layers. Raw conversation logs record what was said, session by session. Automated distillation processes extract the significant content from those logs — themes, facts the user shared, emotional moments, recurring concerns — and compress it into summaries that are cheaper to store and faster to retrieve. These summaries form the working layer of conversational memory.
In SAM, this distillation layer is called SuperMemory. It runs automatically after sessions where significant context was shared, extracting key moments and indexing them for semantic search. Users can see and delete individual memory entries, but do not need to manage the distillation process manually.
Retrieval: when memory enters the conversation
Memory retrieval is not continuous. Injecting memory on every turn would make conversations feel intrusive and noisy — the companion constantly referencing things you said weeks ago would quickly become exhausting. Instead, memory is retrieved selectively: when the current message has high semantic similarity to stored memories, when the user explicitly asks the companion to recall something, or when a recurring theme suggests that prior context would help.
This selective retrieval is called recall gating. A well-tuned recall gate means memory surfaces when it feels natural, not when it feels surveillance-like.
User control over conversational memory
Conversational memory requires explicit user controls to be designed ethically. Users should be able to view what the companion remembers, delete specific memories without deleting others, and delete their full conversation history if they want to start fresh. Memory that users cannot see or control creates a justified sense of unease. See Companion Privacy for the full set of controls.
Frequently asked questions
- What is conversational memory in AI?
- Conversational memory is an external storage system that retains the content of past conversations across sessions, so an AI companion can reference what was said days or weeks ago. It is distinct from a context window, which holds only the current session and is discarded when it ends.
- How is conversational memory different from a long context window?
- A long context window holds more of the current conversation within a single session, but still resets when the session ends. Conversational memory persists across sessions indefinitely — it is stored externally in a database and retrieved as needed, making multi-week and multi-month continuity possible.
- Does an AI companion remember everything I say?
- Typically not verbatim. Conversational memory systems store raw logs but also run distillation processes that extract the significant content — key facts, recurring themes, emotional moments — into summaries. These summaries are what is retrieved during future conversations, not every word of every session.
- When does the companion use conversational memory?
- Memory is retrieved selectively — not on every turn. A recall gate evaluates whether the current message is semantically similar to stored memories, whether the user is explicitly asking the companion to recall something, or whether a recurring theme suggests prior context would help. This prevents memory from feeling intrusive.
- Can I control what an AI companion remembers from our conversations?
- On well-designed platforms, yes. You should be able to view stored memory, delete specific entries, delete full conversation sessions, or clear all memory entirely. SAM provides all three levels of deletion from the Settings panel.