Memory and personalization
Memory is the difference between QRY forgetting your team's fiscal calendar between conversations and QRY quietly remembering it forever. The same memory backend works across all model providers — Claude (native tool), Gemini, OpenAI (custom tools).
This page covers what memory does, how it differs from domain context, and how to manage it without the AI's help when you need to.
How memory differs from domain context
| Domain context | Memory | |
|---|---|---|
| Who sets it | You (or admin), explicitly | The AI, during conversations (you can also edit by hand) |
| When applied | Always, in the system prompt | When the AI judges it relevant |
| Best for | Stable facts: definitions, conventions, business rules | Emergent facts: preferences, recent decisions, one-off context |
| Lifetime | Until you change it | Until you delete it |
| Format | Free-form Markdown | Discrete keyed entries (max 100 KB each) |
A useful rule of thumb: if you'd write it once in onboarding documentation, it's domain context. If it emerged in a chat and you want it to stick, it's memory.
How the AI uses memory
Memory has five operations the model can call as tools:
create— store a new entry.view— retrieve an entry by name.update— string-replace within an existing entry.delete— remove an entry.list— enumerate available entries.
You don't call these — the AI does, autonomously, based on the conversation. You'll see them appear in Processing Details alongside SQL and Python tool calls.
Personal vs. workspace scope
Every memory entry is either personal or workspace-scoped, never both. The database constraint is XOR, which is what keeps the boundary clean.
Personal memory examples
preferred_chart_type: line, not bar, when plotting time series
verbose: false (concise answers, skip the caveats)
preferred_currency: EUR
Workspace memory examples
fiscal_year_start: 2026-07-01
revenue_metric: SUM(net_amount) where status='completed'
deprecated_table: customers (use customers_v2)
internal_email_domain: example.com (always exclude these from
customer metrics)
The XOR boundary is what prevents "Sergio's preferences" from leaking into a teammate's query, and "team's fiscal year" from disappearing if Sergio leaves the workspace.
Telling the AI to remember
You don't need a special syntax. Saying it explicitly works:
Remember that the fiscal year starts July 1. Use that for all subsequent
quarter calculations.
For this workspace, "active customers" means at least one purchase in
the last 90 days. Save that.
The AI decides whether to scope it personal or workspace based on context. If it gets the scope wrong, you can fix it from the Memory page (next section).
Managing memory by hand
Open Settings → Memory under your avatar menu. You see two tabs:
- Personal — entries scoped to you alone.
- Workspace — entries for each workspace you're a member of.
For each entry you can:
- View the full content.
- Edit (text editor, max 100 KB).
- Change scope (personal ↔ workspace).
- Delete.
This is also where to add entries the AI hasn't picked up automatically — sometimes you know upfront that a fact will be referenced often.
Common issues
The AI keeps re-asking something I already told it to remember.
Open the Memory page and check that the entry actually got created. If it didn't, ask again with explicit "save this as memory". If it did, the entry name might not be matching the AI's lookup heuristics — rename it to something more obvious (e.g. revenue_metric instead of metric_def_1).
Memory entry was created with wrong scope. Open the Memory page, find the entry, switch the scope. Existing conversations may have cached the old answer; re-ask for the new scope to apply.
Memory feels stale. The AI doesn't auto-expire entries. If a fact changes (fiscal year, table deprecation), update or delete the entry yourself — the AI uses what's there.
Memory entries leak into a different domain. Personal memory applies to all your conversations regardless of workspace. If you've put workspace-specific facts in personal memory, they'll surface in unrelated conversations. Move them to the right workspace.
See also
- Domain context — for stable, declarative facts that should always apply.
- Shared files and memory — workspace memory in the team-collaboration context.
- Unified Memory Tool reference — full feature reference, including the cross-provider implementation details.