No memory between chats

Between two chats the model retains nothing, and prior context inside one chat reaches the model only when it is supplied again or explicitly referenced, with only the portion that still fits inside the context window actually arriving.

This is how the interfaces are built. Anthropic's API reference describes its messages endpoint as serving single queries or "stateless multi-turn conversations", and OpenAI's guide states that each generation request is independent and stateless, with prior turns visible only when the caller supplies them or points at an earlier response.¹² Anthropic's context-window page gives the mechanism from the other side: each turn's input contains the whole previous conversation plus the new message, and earlier turns survive only because they are sent once more.³ The figure below plays through one chat's turns piling up on the left while a fresh chat on the right stays empty by default, since nothing carries over unless something explicitly supplies or references it.

<div class="wrap">
  <div class="col">
    <div class="hd">One chat, turn by turn</div>
    <div class="b b1">you ask</div>
    <div class="b b2">it answers</div>
    <div class="b b3">you add a detail</div>
    <div class="b b4">it answers again</div>
    <div class="note">Each new turn sends this whole list again</div>
  </div>
  <div class="col">
    <div class="hd">The next chat, just opened</div>
    <div class="blank">empty</div>
    <div class="note">By default, nothing from the chat on the left is in here</div>
  </div>
</div>
<div class="foot">The model itself holds nothing between requests. Memory that crosses chats is information the surrounding application stores; Anthropic's own memory tool happens to keep that as files.</div>
<style>
.wrap { display: flex; gap: 14px; font-family: system-ui, sans-serif; font-size: 14px; }
.col { flex: 1; border: 1px solid currentColor; border-radius: 8px; padding: 10px; min-height: 190px; }
.hd { font-weight: 600; margin-bottom: 8px; }
.b { border-radius: 7px; padding: 6px 9px; margin-bottom: 6px; background: rgba(127,127,127,0.16); opacity: 0; animation: pop 9s infinite; }
.b1 { animation-delay: 0.4s; }
.b2 { animation-delay: 1.6s; }
.b3 { animation-delay: 2.8s; }
.b4 { animation-delay: 4s; }
.blank { border: 1px dashed currentColor; border-radius: 7px; padding: 18px 9px; text-align: center; opacity: 0.6; }
.note { margin-top: 8px; font-size: 13px; opacity: 0.75; }
.foot { margin-top: 12px; font-family: system-ui, sans-serif; font-size: 13px; opacity: 0.8; }
@keyframes pop {
  0%, 4% { opacity: 0; transform: translateY(6px); }
  10%, 82% { opacity: 1; transform: translateY(0); }
  90%, 100% { opacity: 0; transform: translateY(0); }
}
</style>

Figure: one chat's turns pile up because they are resent; a new chat starts empty unless an application layer carries something over.

Three consequences follow inside a single chat. The same page records that this working memory degrades as it fills, a pattern it names context rot, with accuracy and recall falling as the token count grows; it is documented as an effect and no cause is given.³ Chat interfaces may manage a long conversation on a rolling first-in, first-out basis, dropping the beginning. And on Anthropic's API, an input larger than the window is refused outright with an invalid-request error rather than being trimmed.³

Memory across conversations is a separate product feature rather than a property of the model. Anthropic's memory tool stores and retrieves files that persist between sessions, operates on the client side, and keeps that store inside the application rather than inside the model; the instruction it injects tells the model to assume its context may be reset at any moment.⁴ OpenAI documents its consumer memory as optional, inspectable, deletable, and absent from custom GPTs, where each conversation starts fresh.⁵

Being stored is not the same as being remembered. Anthropic's privacy policy says a deleted conversation leaves your history at once and is removed from the back end within thirty days, and that inputs and outputs may be used for training unless you opt out.⁶ OpenAI's platform keeps response objects for thirty days by default.² Storage alone does not put any of that in front of the model; reaching it on your next question takes an application that explicitly supplies or references it.

The boundary: this is a default, not a wall. Layers built around the model do carry information forward, and the exact behaviour of any consumer memory feature is a product decision, so check it against current vendor documentation rather than treating what is claimed here as fixed.