People usually cannot tell what the system actually remembered; the context needs an inspectable view

What it is

A chat log you can scroll is not the context actually used to condition generation. Summaries, retrieved old memories, silently dropped attachments, a persona in the system prompt — none of that is on screen. People can only guess “does it still think I am a beginner,” “is that table still there.” An inspectable context view answers “what sits in front of the model right now,” not another copy of the bubbles.

Memory that cannot be seen cannot be checked, and cannot be changed with a target. People then gamble with a full restatement or a new session. Both are expensive.

Why it happens

Products make “conversation UI” and “model input” two worlds. The UI folds, omits, and pretty-fies for readability; the model input summarises, truncates, and splices retrieved spans for length. The gap is invisible, so the mental model stays at “it sees what I see.” That default happens to hold in a short chat. Once memory, tools, and a system prompt are on, it fails systematically.

Guessing takes two wrong roads. One assumes too much: the uploaded file is still there, and requests are built on a ghost attachment. One assumes too little: a preference already remembered is told again, burning quota and sometimes creating two contradictory versions.

Studying it

Before a real request goes out, have the user draw or list “what you think the system knows right now,” and compare it with what is actually sent to the model. Dependent variables: omissions (present, thought absent), fabrications (thought present, actually absent), and each type’s contribution to later failed requests. Independent variables: whether a context view is shown, grain of the view (summary / source text / typed: persona, memory, attachments, this-session dialogue).

Do not use “does it feel transparent” as the endpoint. Felt transparency can be high while the lists still fail to match. Primary endpoints are list alignment, and whether wrong requests fall after alignment.

Where it stops holding

A stateless, single-turn call that carries nothing hidden beyond a system prompt does not need a dedicated view; the log is the context. Showing the full system prompt may leak safety policy or other people’s data; the view has to be graded: people have a right to see memories about themselves and this-session material, not necessarily every system instruction. A raw view of extreme-length context is itself unreadable; typed lists plus “why this is here” beat pasting a hundred thousand tokens.

Applying it

• Provide a panel “what it sees right now,” split by source: this session, cross-session memory, attachments, system assumptions about the person. Each row expands to enough source or summary to judge true or false.
• Objects the person thinks are in, and that are actually out (expired attachments, details dropped by a summary), must be marked out, not omitted. Omission is read as still in.
• The view must open from the current session, not only from a Memory page under account settings. The decision happens mid-dialogue.
• Check: have people list three things the system knows without opening the panel, then open it and reconcile. Fabrications and omissions should both drop. Then send a request that depends on an attachment they “thought was still there” — if the panel marks the attachment out, that request should be withdrawn before send.

Related

• Same group: L2.14.1 Context buys continuity, and it also lets an early misunderstanding keep shaping later output · L2.14.3 Clearing needs an explicit act; opening a new session is only an implicit convention · L2.14.4 Memory across sessions hardens a one-shot preference into a standing setting · L2.14.5 When overlong context is truncated, what the user sees is unexplained capability loss
• Nearby: L5.05 The Moderation Principle of Transparency · L2.07 Prompt History and Reuse · L6.10 Turning Personalization Off and Resetting It
• Search terms: inspectable context view · model input transparency · what the model sees