When you say “keep your answers brief,” you may want that preference to last. Saying “I'm tired today” does not necessarily invite a permanent judgment about you.
ISEA stores lasting information separately from chats. All three agents can use the same account's memories and should use the updated information after you correct or delete an entry.
Where memories come from
A stable preference, a lasting fact, or something you explicitly ask the assistant to remember can become a memory. To save it, the tool must quote your current message. The server keeps that source with the entry.
Temporary emotions, quotations from other people, inferred identities in images, and credentials should not become lasting personal facts. Sensitive facts require an explicit request. Meals, schedules, and expenses belong in business records rather than vague memory entries.
Having a source does not guarantee that the model understood it. Memories can still be wrong or incomplete. You can ask to see them and correct what was saved.
Reading details when needed
At the start of a turn, the assistant sees memory categories and the current version. It reads the details when needed, rather than adding everything to every conversation. The version lets it check whether it has the latest information.
If you used to prefer morning exercise but explicitly switch to evenings, later tasks should use the updated preference rather than revive the old one from a summary.
Corrections, deletion, and stale writes
Memory updates carry version information. When two tasks start from an older version, the server refuses a stale write that would overwrite a newer update. The task must read current information again.
Deletion advances the version too, preventing an older request from recreating the deleted entry. Temporary background can have an expiration time and stops appearing as valid memory after it expires.
A running task may already have read the old information, so check its result after a correction. Updating memory also cannot reverse actions that have already finished.
Memory is different from history
Conversation history preserves exchanges. Summaries help a long conversation fit a bounded context. Persistent memory retains selected preferences and facts. Business records preserve events.
Summaries can omit details, and memory is not a complete history index. “How much did I spend last month?” belongs to the ledger. “How do I want you to respond?” may belong to communication preferences.
Deleting a memory does not automatically delete the chats or business records where that information appeared. Later tasks should use current valid information, while deletion scopes remain separate.
Expressing what you want retained
You can say “remember that I prefer brief answers,” or “correction: for the next month, I'm only available in the evening.” When background is no longer relevant, explicitly ask for it to be removed.
After a correction, ask the assistant to read the current memory so you can check it. In later conversations, notice whether it uses that updated information.