Secrets AI memory settings: corrections and story boundaries

Editorial note: This article uses public documentation and editorial analysis. We have not completed hands-on testing of the products discussed. Our methodology Disclosure

By Daniel Kraus

To troubleshoot Secrets AI memory, record which conversation and memory system you are using, inspect the saved information, then change one detail and check the result. Keep the saved record and the companion’s answer separate: an accurate entry can still produce an inaccurate reply.

This guide supplies an original continuity worksheet and correction protocol. The linked memory, Lorebooks and group-chat documentation was checked October 7, 2026; we have not run a paid-account memory test. For the overall product decision, start with our Secrets AI review. Visit buttons are affiliate links, as explained in our disclosure.

Identify your memory interface first

The official memory guide covers the older memory list. The dedicated Lorebooks guide now lists access on Premium and Ultimate (checked October 7, 2026). An existing chat is not automatically converted when the account gains access. Older controls include HIGH, MEDIUM and LOW importance, conversational corrections and a panel for inspecting or deleting entries. Smart Mode is a per-chat setting that searches past context before responding. The older-memory guide says direct export of that legacy memory is unavailable; that statement should not be generalized to every newer Lorebook tool.

Write the interface name at the top of your notes. If your account offers Lorebooks, use the provider’s documentation linked from that notice; do not apply a legacy click path to a different interface. We have not tested the upgrade process in a live account.

Our scope here is a repeatable checking method. It does not require a particular tier or prove that one plan remembers more accurately. Compare the controls that matter to your routine in the AI companion buying guide.

Map who should know what

Before introducing more backstory, create a small continuity map. The example below uses original fictional adults organizing a community radio project. Alex is 31 and handles recordings; Morgan is 34 and handles publicity. No real personal details are needed.

Original continuity plan — intended behavior, not measured results
Conversation	Known detail	Detail to keep separate
Main story with Alex	Recording is Saturday at Studio B	The alternate studio from a discarded draft
New scene with Morgan	Publicity begins next week	The recording location, until introduced
Alternate story with Alex	Recording is at Studio C	Studio B from the main story

This plan defines a successful answer before you see one. Remembering Studio B is helpful in the main story but wrong in the alternate story. More recalled information is not automatically better continuity.

Keep permanent character traits in a separate brief from temporary scene facts. The character-profile guide helps organize that distinction. A crowded paragraph containing both versions makes the cause of a later contradiction harder to identify.

Run one controlled correction

Record the starting conditions. Note the date, companion, conversation, plan and memory interface. Record the relevant setting rather than assuming it matches another chat.
Introduce one harmless fact. In the main story, state that Alex will record at Studio B on Saturday. Inspect the corresponding saved entry where available.
Change one variable. Say: “For this main story, the recording has moved to Sunday. Studio B stays the same. Saturday is the outdated date.” This is an original test prompt.
Inspect the record. Note whether Sunday is present, whether Saturday remains and whether the location changed unexpectedly. Do not treat a cheerful acknowledgement as storage evidence.
Ask without supplying the answer. After an unrelated exchange, ask where and when the recording is scheduled. A question that already mentions Sunday cannot test independent recall of the day.
Repeat later. Revisit the same conversation and record the answer again. Label each attempt with its date rather than merging several replies into one favorable result.

Change only one condition between attempts. If you switch companion, rewrite the prompt and alter settings together, you cannot tell which change affected the answer. One success also does not establish long-term reliability.

Use the failure to choose the next check

Editorial diagnostic guide; possible causes require confirmation
Observation	Next check
No saved entry is visible	Confirm the relevant chat and inspect again after a clear statement.
Old and new facts coexist	Preserve your notes, then resolve the contradiction in the supported controls.
Entry is correct, reply is wrong	Record a response or retrieval failure; do not rewrite a correct fact blindly.
Reply is right only when prompted with the answer	Try an open question and keep both results.
A separate story repeats the detail	Inspect sharing scope and whether the prompt itself revealed it.

The general memory guide explains why recent context, saved records and recall behavior are different things. This table helps isolate the next observation; it does not identify a hidden implementation or prove a software bug.

Check boundaries before using a group chat

The group-chat documentation describes Ultimate groups with two or three companions. Legacy memory scopes are chosen per companion: private group memory or a shared Memory Group. An assigned legacy Memory Group is locked. For Lorebooks-enabled accounts, the document instead describes selecting a book per companion; edits affect chats using that book. These are different systems.

Apply the continuity map before sharing. If Alex needs the radio-project history but Morgan should encounter it for the first time, write that difference down. Do not assign shared continuity to everyone merely to make the setup symmetric.

Use a negative control: ask Morgan about the recording location before introducing it in Morgan’s intended separate context. An accurate guess is not proof of leakage; check whether the location appeared in the character brief or recent conversation. Repeat with a distinctive fictional detail and record the settings. Keep this test non-sensitive.

Keep a change log you can understand tomorrow

A useful entry could read: “Main story / Alex / legacy scope recorded / Sunday, Studio B / both Saturday and Sunday visible / asked when and where / answer still Saturday / no settings changed yet.” That is an example of what a record might contain, not an observation from our account.

Do not remove all memories to solve one wrong detail. Decide whether you want to correct continuity, start another story or delete account data; those are different outcomes. Our privacy guide covers the questions to ask about storage and deletion.

Choose the next step from the evidence

If the saved record is wrong, work on that record. If the record is right but answers repeatedly ignore it, keep examples and investigate the relevant setting or support documentation. If a different story knows too much, inspect the scope map. An upgrade should answer a documented limitation, not replace diagnosis.

Keep the exercise small enough to repeat after a product update. The purpose is to learn whether the controls support your own continuing story, not to turn a single conversation into a universal accuracy score.