AI Memory
AI memory is the set of mechanisms by which an AI companion stores, organises, and retrieves information from past conversations so that future interactions have continuity and context. Without memory, every conversation starts from scratch. With it, a companion builds a genuine relationship across time.
The three memory layers
Most memory-capable AI companions use three distinct layers, each serving a different purpose. The session context window holds the current conversation — everything said so far in this exchange. It is the AI's immediate working memory, typically limited to a fixed number of recent messages.
The curated memory layer (called the Heartline in SAM) is a store of specific reflections, moments, and facts that the user writes and controls directly. These entries are explicitly remembered and can be edited or deleted at any time.
The distilled summary layer (called SuperMemory in SAM) automatically extracts and stores key themes and context from past conversations. These summaries are indexed using vector embeddings and retrieved when they would be relevant to the current turn — without requiring the user to manage them manually.
How retrieval works: embeddings and recall gating
Storing memory is only half the problem. The other half is retrieving it at the right moment — not on every turn (which would be noisy and overwhelming) but only when it would genuinely help the response. This is called recall gating.
When a user sends a message, the companion converts it to a vector embedding — a mathematical representation of its meaning — and compares it against stored memory embeddings. Entries with high semantic similarity are candidates for injection. A recall coordinator then decides whether to actually inject them, based on signals like explicit recall intent ("do you remember when I said…"), repeated themes, and emotional context.
See Conversational Memory for a deeper look at how session and long-term memory interact.
What memory enables in a companion relationship
Memory is the mechanism that turns a series of separate conversations into a relationship. A companion without memory is always meeting you for the first time. A companion with memory knows your name, recalls what you were going through last month, notices when you return to a theme you discussed weeks ago, and can ask a follow-up question that acknowledges what came before.
This continuity changes the quality of the interaction fundamentally. Users do not have to re-explain their situation every session. The companion does not have to ask basic context questions it already knows the answer to. The conversation can start in the middle, exactly where it should.
The difference between memory and training
A common misconception: that an AI companion with memory has been trained on your conversations and the underlying model has changed. This is not how it works. The model itself remains stateless and is not retrained on individual user data. Memory is stored externally — in a database — and injected into the prompt on relevant turns. The model sees your past context in the same way it would see any text: as input, not as learned weights.
This distinction matters for privacy. Deleting your memory (your Heartline entries, your conversation history) removes it from the retrieval pool. The model has not been changed by your data, so there is nothing to "untrain".
Frequently asked questions
- What is AI memory?
- AI memory is the mechanism by which an AI companion stores information from past conversations and retrieves it in future ones. Without it, every conversation resets. With it, the companion builds continuity — remembering your name, recurring themes, past discussions, and what matters to you.
- How does AI memory actually work technically?
- Past conversations are converted into vector embeddings — mathematical representations of meaning — and stored externally. When you send a message, the system compares it against stored embeddings and retrieves the most semantically relevant memories. A recall coordinator then decides whether to inject them into the current prompt.
- Is AI memory the same as training the model on my data?
- No. The underlying language model is not retrained on your conversations. Memory is stored separately and injected as text into the prompt when relevant. Deleting your memory removes it from the retrieval pool entirely — there is nothing to untrain because the model has not changed.
- Can I delete what an AI companion remembers about me?
- Yes, on well-designed platforms. SAM offers three deletion options: delete individual conversations, delete individual Heartline memory entries, or delete your full account and all associated data. Deletions take effect in the retrieval pool within a few hours as the index updates.
- What is recall gating?
- Recall gating is the process of deciding whether to inject long-term memory into a given turn at all. Not every turn benefits from memory injection — injecting it indiscriminately would make conversations feel intrusive and noisy. A recall coordinator evaluates intent, emotional context, and thematic signals before deciding whether to surface stored memories.