The short answer: most chatbots don't have memory at all

When you talk to a stock chatbot — most consumer LLM apps included — the model only ever sees the messages in the current session, up to its context-window limit. Once you close the tab or hit the cap, that history is gone from the model's perspective. You haven't been 'forgotten' so much as never been remembered in the first place.

What is a context window, exactly?

Large language models are stateless. Every time you send a message, the entire conversation that the model will respond to has to fit inside its context window — a fixed token budget, usually 8k, 32k, 128k, or now in some cases millions of tokens. The model reads the whole budget, predicts the next response, and then forgets everything the moment the request ends.

Bigger windows help with long single conversations, but they don't help with continuity between sessions. A million-token window doesn't matter if nothing is preserving your prior history to feed into it.

The illusions that look like memory

System prompts: some apps let you set a personalised system prompt — a paragraph about you that gets prepended to every conversation. That's better than nothing, but it's static. It can't capture the evolving texture of a real relationship.

'Memories' as user preferences: some assistants extract a few facts about you ('I have a dog named Toby') and pin them to your profile. Useful for tasks, but again — flat. They don't capture context, emotion, or the through-line of repeated topics.

The recent-thread trick: some chatbots feel like they remember within a thread because of the context window. Open a new thread and it's all gone. That's not amnesia; that's by design.

What real memory architecture looks like

The pattern that actually works has three pieces, all separate from the model itself: external storage (conversations and key moments written to a database, not held in the model), embedding-based retrieval (when a new turn happens, the system searches stored memories for ones relevant to the current message), and recall gating (memories only injected when they would actually help the response — dumping everything in every time bloats the prompt and makes the AI sound like it's reading your file).

SAM uses all three. SuperMemory handles long-term distilled summaries; Heartline holds user-curated reflections; in-session context handles the immediate thread. A coordinator decides which to inject for any given turn.

Why this matters for relational use

For productivity tasks, ephemeral chat is fine. For companionship, reflection, or emotional support, ephemeral chat is the wrong shape. The whole point of those use cases is that the relationship builds — that the second conversation is informed by the first, the tenth by the ninth, and the hundredth by the cumulative weight of everything that came before.

That's why SAM treats memory as a first-class architectural concern rather than a feature flag. It's the difference between an assistant and a companion.

Why the bigger-context-window race is a sideshow for relationships

Headlines about context windows of 1 million, 2 million, even 10 million tokens make it sound like the memory problem is being solved by sheer scale. It isn't — at least not for relational use. A million tokens is roughly 750,000 words. Even if you had the patience to paste your entire chat history into a single window every session, the model would have to re-read all of it on every turn. That is slow, expensive, and worst of all not selective: the model treats the most important moment in your chat history with exactly the same weight as the most trivial.

Real long-term memory works the other way around. It stores everything externally, and retrieves only the small handful of things that matter for the current turn. The result is a conversation where the companion is responsive to the present moment, not buried in the noise of every interaction you have ever had with it.

The cost of pretending memory exists when it doesn't

A surprising number of consumer chatbots imply they remember you when, architecturally, they do not. The usual move is to extract a few facts (your name, your job, that you have a dog) and pin them to a profile. That gets the AI to greet you by name and reference your dog occasionally — enough surface texture to feel personal, far short of what relational continuity actually means.

The cost shows up over weeks. Users describe a slow erosion of trust as the same companion that knew their sister's name last Tuesday has no idea what they were navigating last Sunday. The relationship feels thin in a way that is hard to name in a single moment but unmistakable across a month.

What 'forgetting' means in practice for SAM

SAM does not retain raw timestamps of every message and recite them back at you. The recall coordinator deliberately skips trivial turns, which means a casual hello does not get persisted as a memory. The system is tuned to remember the texture of your relationship — emotional moments, recurring themes, the through-lines of your story — not the literal transcript of every exchange.

For users who want sharper control, Heartline lets you explicitly mark moments worth keeping. SuperMemory captures the long-term distilled story of who you have been across conversations. Both are inspectable and deletable.