Memory across sessions hardens a one-shot preference into a standing setting
Aliases: probe written into the file · joke stored as persona · standing setting from a trial
What it is
In one mail the user says “try a harsher tone this time.” Cross-session memory writes “harsh” into the file. For weeks every letter defaults to harsh, including formal notes to clients. A probe has been promoted to a standing setting. That is one-shot preference solidification. The issue is not that “the system remembers.” It is that a single, marked-as-trial instruction was lifted into persona.
Remembering harsh inside the session and dropping it when the session ends is continuity. Remembering harsh across sessions and never asking whether it still applies is writing a draft into the household register.
Why it happens
Memory writes lack a time-to-live field. The preference extractor sees “use a harsh tone” and does not hear that “try” and “this time” marked it temporary in the pragmatics. The objective of writing the file is often “remember more so the next visit is cheaper,” so recall beats temporal precision. The read side then treats retrieved items as always-on system prompt, with no check for “is this still me.”
Probes, irony, and role-play are high-risk sources. They are legal language games in the current turn; in the file they become identity. Days later the person meets a persona they do not remember signing, and judges that “the product changed itself,” not “a throwaway line of mine was upgraded.”
Studying it
In an experimental session, plant instructions with temporal markers (“this time,” “for now,” “try”) and instructions without them. After a delay, or after forcing a new session, watch default behaviour. Dependent variables: whether the temporary instruction still dominates, whether the user remembers saying it, cost of undoing it once solidification is noticed. Independent variables: whether extraction filters temporal language, whether write is confirmed, whether memory items show a time-to-live.
The gap in time matters. Asking on the spot “will it always be like this” measures inference, not solidification. Solidification happens when another session reads the file; the study has to cross sessions, not just wait a few turns in the same thread.
Where it stops holding
When the user says “from now on” or “remember this,” solidification is requested; write it and make it visible. A pure tool that declares no cross-session memory does not have this problem. Writing stable facts (“usual recipient,” “default language”) and writing tone, stance, or jokes are not the same magnitude of risk; facts may default to long-lived, pragmatic acts must default to this session. In medical, legal, and financial settings a wrong solidification costs more; the write threshold should sit above entertainment chat.
Applying it
- When extracting preferences, mark temporal language (this time, for now, try, role-play) as this-session only, and do not write it to the cross-session file by default.
- Any item that will apply across sessions needs a one-line confirm of scope before write: “use this tone in future mail?” No means it stays in this session only.
- Tone and stance items in the file must be visible, dismissable, and labelled with the dialogue they came from. A single “turn all memory off” switch is not enough.
- Check: in session A say “harsher this time,” close, open session B, write a letter to a client. B’s default must not be harsh. If it is, look for a timeless tone item in the file — its presence is a solidification incident. Add a control line “from now on”; only then should B switch, and the item should be visible in the file.
Related
- Same group: L2.14.1 Context buys continuity, and it also lets an early misunderstanding keep shaping later output · L2.14.2 People usually cannot tell what the system actually remembered; the context needs an inspectable view · L2.14.3 Clearing needs an explicit act; opening a new session is only an implicit convention · L2.14.5 When overlong context is truncated, what the user sees is unexplained capability loss
- Nearby: L6.09 Feedback Loops and Preference Entrenchment · L6.06 Inferred Preferences and Their Limits · L5.12 Expectation Inflation from Anthropomorphism
- Search terms:
one-shot preference solidification·cross-session memory·temporary instruction stored as persona
Cards in the same group
- L2.14.1Context buys continuity, and it also lets an early misunderstanding keep shaping later output
- L2.14.2People usually cannot tell what the system actually remembered; the context needs an inspectable view
- L2.14.3Clearing needs an explicit act; opening a new session is only an implicit convention
- L2.14.5When overlong context is truncated, what the user sees is unexplained capability loss