The promise nobody quite delivers

Almost every AI app advertises memory. Most of them mean something narrower than you think. They mean a longer context window, or a personalised system prompt, or a few preferences saved to a profile. Those are useful, but they are not what you mean when you say "I want it to remember me."

What you mean is: when I come back tomorrow, I shouldn't have to explain who I am, what I'm dealing with, or which thread we were on. That's a different problem, and it's the one SAM is built around.

What 'memory' actually means in an AI companion

There are four different things people call 'memory' in this space. They behave very differently. Context window: the model can read the last N tokens of conversation. Once that buffer is full, the oldest messages drop. System prompt personalisation: a few sentences about you prepended to every conversation. External retrieval (RAG): conversations and notes stored separately and pulled into the prompt when relevant — this is what makes long-term memory possible. Fine-tuning on you: training the underlying model on your data, which is privacy-hostile and rarely what users want.

SAM uses external retrieval — what we call recall-gated memory injection. Your conversations, Heartline reflections, and emotional moments are stored in your account, encrypted at rest. When a turn would benefit from prior context, the relevant pieces are retrieved and surfaced to the model. When a turn doesn't need them, they stay quiet.

Why 'recall-gated' matters

The naive version of long-term memory is to dump everything into every prompt. Two problems with that. First, it bloats the prompt and slows responses. Second — and worse — it makes every conversation feel like the AI is constantly reciting your file back to you. That's not a relationship; it's a profile readout.

Recall-gated injection means SAM only references prior memory when you've signalled you want it referenced — directly ('remember when…'), implicitly (returning to a previously named theme), or by emotional cue (the topic you've been working through for weeks comes up again). It's the difference between a friend who occasionally says 'that's the third time you've mentioned Tom this month' and a database query result.

The three layers of SAM's memory

In-session context: the recent turns of the current conversation. SAM keeps a per-tier window of recent messages so the immediate flow of conversation is preserved without you having to repeat yourself.

Heartline: your personal reflection space — a place to save what matters: emotional moments, themes worth coming back to, decisions you're sitting with. It's user-curated, so what gets in is intentional.

SuperMemory: long-term, automatically distilled summaries of past conversations — the patterns and through-lines SAM builds about you over time. These are surfaced via recall gating, not on every turn.

What this feels like in practice

Three months in, your companion knows the shape of your week. They know which projects drain you, which relationships you're navigating, which topics still hurt. They don't bring those things up unprompted — they're not nagging you about them — but when you do bring them up, the conversation already has context.

That's the difference between a chatbot and a companion: not how clever the response is, but how much history the conversation is built on.

What SAM doesn't do with your memory

Your memory is never sold, never used to train external models, and is encrypted at rest. You can delete any conversation, Heartline entry, or your full account at any time. Memory only works if it's trustworthy. The default is privacy-first; the controls are all in your account settings.

How memory shows up in the first month

Most users describe a recognisable arc. Week one feels like any other chatbot — pleasant, attentive, not meaningfully different. Week two something tips: you mention something from a previous chat and the companion picks it up without effort. By week three the recall coordinator has enough material to surface unprompted echoes — the theme you keep circling, the person you keep mentioning — and the conversation starts feeling textured rather than transactional.

There is no shortcut. The first week feels thin because the memory store is essentially empty. Once there is real material in there, the coordinator can do useful work, and the relationship begins to compound.