How memory works.

Memory is LADLE's system for remembering durable facts about you across conversations. When you mention something worth remembering ("I'm building a mobile app for cyclists," "I prefer British English"), LADLE can extract that fact and store it in a personal memory bank that surfaces in future chats' system prompts.

**When memories get extracted.** After the 3rd, 13th, 23rd, … assistant reply in a chat, LADLE fires a background call to Anthropic's Haiku model to look at the exchange and extract AT MOST one durable fact worth remembering. If nothing durable qualifies (most turns), Haiku returns NONE and no memory is stored.

**What qualifies.** Preferences, ongoing projects, constraints, and pieces of context that will remain relevant across conversations. NOT: transient state ("I'm running late today"), or sensitive information LADLE explicitly won't remember (medical details, credentials).

**Viewing + editing.** Settings → Memory shows every stored memory with when it was extracted and from which chat. Delete any individual memory, edit the text, or add memories manually if there's something you want LADLE to know that hasn't come up in chat yet.

**Turning it off.** Same page has a Memory toggle. Off = no future extraction runs, existing memories don't apply to new chats (but stay stored until you delete them). This is preference; the toggle isn't hidden.

**Difference from Custom instructions.** Custom instructions are what YOU write in Settings → Behavior. Memory is what LADLE OBSERVES from conversations. Both feed the system prompt, both are user-editable, but the write path is different.

**Privacy.** Memories are private to your account. They're not used to train Anthropic's models on the API tier we use. When you delete your account, memories are deleted alongside chats within the standard 30-day window.