A well-tended memory list.

LADLE's memory system remembers up to about 50 facts about you across your conversations. Every memory is visible at Settings → Memory. You can edit, disable, or delete any of them. The assistant will never announce that it remembers something.

Turned on and left alone, memory will accumulate whatever the background haiku-based extractor deems worth saving after every tenth assistant reply. Some of what it saves will be genuinely useful ambient context ("prefers Go over Rust"). Some will be stale within a week ("currently reading Piranesi"). Some will be too specific ("named the current side project 'Otter'"). This is the whole reason the list is visible: so you can curate it.

Memory vs custom instructions

The two are easy to confuse. The distinction:

**Custom instructions** are stable rules about how you want the assistant to behave. They belong in Settings → Behavior and change infrequently (quarterly at most). Example: "Prefer British spelling."

**Memories** are ambient facts about you the assistant can quietly use. They're extracted opportunistically and can be added/removed one-by-one. Example: "Grew up in Auckland, moved to Berlin in 2019."

If you find yourself putting instructions in memory, move them to custom instructions. If you find yourself putting stable facts in custom instructions, move them to memory. The line is: instructions tell the assistant what to do; memories tell it about you.

What good memories look like

- **Durable identity facts.** "Works as a designer at a fintech." "Lives in Melbourne." "Reads mostly nonfiction." - **Stable preferences.** "Prefers concise replies unless asked for depth." "Doesn't use emoji in professional writing." "Runs macOS with a mechanical keyboard." - **Ongoing projects.** "Currently building an AI onboarding tool called Otter." "Writing a book on urban design, targeting Fall 2026." - **Domain expertise you don't want re-explained.** "Has ~10 years of backend engineering experience." "Was a working journalist for six years."

What bad memories look like

- **Ephemera.** "Watched Dune 2 last night." — will be false in a week. - **Overly-specific details from one conversation.** "Was frustrated with the Zoom meeting at 4pm." — was true for one hour. - **Anything that reads as a note-to-self rather than an ambient fact.** "Should look into fixing the CI build." — belongs in your task manager, not the model's ambient context. - **Duplicates.** Memory does dedupe extracted facts against the seeded role sentence, but general dedupe is a known gap. If you see near-duplicates, delete the older one.

The seeded onboarding memory

If you completed the Setting the Table onboarding, one memory is already seeded: your role sentence, marked with source='onboarding'. It looks like "The user works as a designer." Leave it or edit it — the source label makes it easy to find.

Curation cadence

Once a month is enough. Open Settings → Memory. Read the list top-to-bottom. For each:

- **Still true, still useful?** Leave it. - **Was true, no longer applicable?** Delete or edit. - **Too specific?** Delete. - **Actually a preference you want to enforce always?** Move to Settings → Behavior.

The list should stay roughly at 20-40 items, tuned to who you are now. If it's over 50, the tail is probably not helping — the composed system prompt is capped at ~2000 characters of memory anyway, so the least-recently-updated memories are getting silently truncated.

Turning it off

Memory is a single toggle in Settings → Memory. Off = no memories load into the system prompt AND no new memories are extracted (the background haiku call skips). Turning it back on later leaves your existing list intact; extraction resumes.

Some people prefer to leave it off entirely and set every stable fact as a custom instruction. That works. It gives you tighter control at the cost of the ambient-context ease. Both configurations are legitimate.

What the assistant does with your memories

Silently informs the reply. If you ask "recommend a book" and the assistant knows you read mostly nonfiction, you'll get nonfiction recommendations without having to say "I read mostly nonfiction." If you ask for editing feedback and the assistant knows you're a former working journalist, it'll adjust the depth of its critique accordingly.

What the assistant does NOT do: name the memory, cite the memory, use the phrase "I recall from our conversation." If you see the assistant doing that, it's a bug — email help@ladle.chat.

The privacy note

Memories are user-scoped rows in your account. RLS ensures no one else can read them. They're included in your data export. They're deleted (irreversibly) if you delete your account. They're not shared with any third party. They're not part of the training data for any future model.

They exist purely to make the assistant feel like it knows you a little, in the way a friend who's known you for a year knows you a little, without needing to be re-briefed every conversation.