Account memory

• Standing preferences — tone, format, length, language.
• Recurring constraints you always want applied, such as a reporting currency.
• Facts about your own role, business or priorities that stay true across sessions.

Anything that should be visible to teammates belongs in workspace memory instead. That boundary is enforced, not advisory.

Account memory is scored alongside conversation context and workspace memory during context assembly. It participates in ranking, so it informs an answer but never silently overrides an explicit instruction in the current question.

1. 1State the correction explicitly in Ask, for example "remember that my reporting currency is USD, not LKR".
2. 2Confirm the updated behaviour in a later conversation.
3. 3Ask for removal of a fact you no longer want retained, and verify it is no longer applied.

If a stored fact is wrong and keeps appearing, correct it in a new conversation rather than trying to override it repeatedly — repeated overrides do not delete the original.

Good account memory is small, stable and reusable: a reporting currency, a preferred output format, the role you hold, the market you sell in. Bad account memory is anything that changes weekly, anything a connected service can provide more reliably, and anything that actually belongs to the team.

• Store the rule, not the instance — "always show monthly totals" beats a list of individual months.
• Store corrections as corrections, so the wrong version stops competing with the right one.
• Do not store anything a connection can already answer with fresh, sourced data.

• Never visible to another user, including teammates in the same workspace.
• Never included in public documentation, sitemaps or machine-readable exports.
• Never used to bypass a permission: memory cannot authorise access to data.