LADLE has three layers of instructions but no automatic memory. This is a deliberate product decision — automatic memory sounds convenient but produces surprising, hard-to-debug behaviour. Every context the model uses is either in the current chat, in the project the chat sits inside, or in your global custom instructions. All three are visible and editable.
The three instruction layers
Instructions cascade from global → project → chat. More-specific layers override less-specific ones for the current chat.
Why no automatic memory
Products that auto-remember across chats (fact extraction, user modeling, silent context injection) trade transparency for convenience. Users can't easily see what the model 'knows,' can't audit its assumptions, and can't reset without losing everything. LADLE's stance: if the model needs to know something, put it in one of the three explicit instruction layers. Everything the model uses is inspectable and editable — no hidden state.
How to use the instruction layers well
Global instructions: identity + preferences that follow you everywhere (your name, your work role, your tone preferences, formats you want by default). Project instructions: context specific to the project's domain (this is a legal matter for X, this codebase uses Y conventions). Chat instructions: one-off overrides (respond only in JSON for this session, ignore project voice for this drafting exercise). Keep each layer short — instructions consume tokens on every message.
Instructions and token usage
Each active instruction layer is prepended to your prompt on every message. A 500-word global instruction plus a 1,000-word project instruction adds ~2,000 input tokens per message. This is usually negligible for base-tier usage. If you're rate-limited, check how much your instructions add via Settings → Usage (they're broken out).
Resetting and editing
Each layer is text you can edit at any time; changes take effect on the next message. There is no 'forget everything you know about me' — because there's nothing hidden to forget. To restart from a blank slate: clear the relevant instruction layer, or start a new chat outside any project.
Why not just use the 200K context for memory
You can. Paste a long history summary into the chat as context. LADLE won't retain it after the chat ends — that's the honesty part. If you want a memory that persists across chats: put a summary of what matters in the project or global instructions. It stays visible and you can edit it.