# Context & Memory Producer Pal remembers things about your music in three layers. Each one is included in every conversation, and each answers a different question: | Layer | Answers | Stored | | ------------------- | ------------------------------ | ---------------------------------------------------------------- | | **Project context** | What is _this_ Live Set? | In this project's Max for Live device — it travels with the Set | | **Global context** | What do you always want? | On your computer, in `~/.producer-pal/context.md` | | **Memory** | What has AI learned about you? | On your computer, one file per fact in `~/.producer-pal/memory/` | `~` is your home folder, so `~/.producer-pal` is: - **macOS** — `/Users/you/.producer-pal` - **Windows** — `C:\Users\you\.producer-pal` That folder holds the files behind every layer, so you can open it to edit or back them up directly. You write project and global context by hand. Memory is AI's own notebook: it records facts as you work, and you can edit them. AI can edit all three with the [Context tool](/features#ppal-context). All three reach every client — the built-in [Chat UI](/guide/chat-ui), external MCP clients like Claude Desktop, and the [REST API](/guide/rest-api) — because they are attached to the result when AI connects. ## The context editor The editor has a tab per layer, plus two tabs for customizing how AI is prompted. Open it from: - The **Context** button in the [Chat UI](/guide/chat-ui) header - The **Open Editor** button on the device's [Context tab](/guide/device#context-tab) - The **Edit Context** link beside the Context tool in the Chat UI's Tools settings Every tab shares the same chrome: edits **auto-save** as you type (the header shows "Saved"), a size readout gives the character count and approximate token cost, and the import (⬆) / export (⬇) buttons read and write markdown files — you can also drag a `.md` file onto the editor. AI and the editor stay in sync: if AI writes while you have a tab open, a banner offers to reload. ## Project Project tab Notes about the Live Set you're working on: its genre and tempo, how the arrangement is laid out, which track does what, and the rules you want followed. Project context is saved in this project's Producer Pal device, so it travels with the Live Set and is gone if you delete the device. This is the same text as the device's [Context tab](/guide/device#context-tab) — the device shows it in a small box, the editor gives it room. Keep it about the project, not about you. "Kick stays four-on-the-floor" belongs here; "I always work in A minor" belongs in Global. ## Global Global tab Notes that apply to everything you make: the genres and tempos you work in, your production habits, conventions you want AI to respect, and things you never want it to do. Global context is stored on your computer, so it follows you from project to project — a new Live Set starts with your preferences already in place. ## Instructions Instructions tab The system prompt for the built-in Chat UI. It ships with a default (shown read-only, with its size) that tells AI to act as a music-composition assistant and to use its tools rather than asking you for details it could look up. Press **Customize** to fork the default into an editable copy, and the trash button to reset back to the default. Only the built-in Chat UI uses this — external apps like Claude Desktop or Claude Code bring their own system prompt. ::: warning This replaces, not appends Custom instructions replace the default entirely. If you want AI to keep behaving like Producer Pal's assistant, start from the default (Customize forks it for you) rather than from a blank page. To _add_ a preference, use Global context instead — that's what it's for. ::: ## Skills Skills tab The **Producer Pal Skills** are the instructions AI receives when it connects — how to write notes, edit clips, build instruments, and search your library. The Skills tab lets you override any fragment of them. The dropdown picks a fragment; each shows read-only with a **Customize** button until you fork it. Customized fragments are marked ✎ in the dropdown, and ⚠ if the built-in changed in a newer Producer Pal release after you forked it (your override still applies — the mark just tells you it's now behind). Once you customize a fragment, the editor shows your override beside the built-in so you can see what you changed: A customized skills fragment - **Your override** (left) is editable and auto-saves - **Default** (right) is the built-in, with **Copy** to grab its text and **Hide** to collapse back to one column - The trash button beside "Default" resets the fragment, deleting your override **Preview** shows what AI actually receives — the whole assembled document for a given notation and model size, with your overrides applied. The customized library fragment above appears here in place, right where it gets pulled in: Skills preview The ★ badge marks the combination your current settings use, and the size readout is why trimming matters: everything here is sent on every conversation. **Source** switches back to editing. See [Customizing Skills](/guide/customizing-skills) for how the fragments fit together, and how to drop whole areas you never use. ## Memory Memory tab Memory is AI's own notebook about you, kept across projects. Where global context is a blob you write, memory is a list of small facts AI writes as it learns them ("you never want hats hard-quantized"), stored one file per fact. Each entry has three parts: - **Name** — a short slug, and the filename in `~/.producer-pal/memory/` - **Description** — a one-line recall hook - **Memory** — the fact itself Only the **names and descriptions** are sent to AI on every conversation, as an index. It reads a full entry only when the description suggests it's relevant. That's what keeps a growing memory cheap: a hundred facts cost a hundred lines, not a hundred bodies. ::: tip Write descriptions as hooks, not titles The description is the _only_ thing AI always sees, so it has to say when the memory matters. "Dislikes rigidly quantized hi-hats; wants swing" earns a lookup. "Hi-hat stuff" does not. ::: Use **New memory** to add an entry yourself, click one to edit it, and the trash button to delete it. Renaming an entry renames its file. ::: details Memory in small model mode [Small model mode](/features#small-model-mode) turns memory off: the index isn't sent, and AI can only read and write project and global context. Smaller models do better with fewer moving parts, and memory's load-on-demand step is one they tend to skip. ::: ## What AI sees when it connects On connect, the three layers are attached to the result in order: your project context, then your global context, then the memory index. Skills and the context layers are all sent up front (as is your custom system instruction, in the built-in Chat UI), so if a conversation starts before you make an edit, start a **new conversation** for the change to take effect. To stop AI from reading a layer, empty it. To stop AI from _writing_ to any of them, turn off the Context tool (the **Context** checkbox under the Chat UI's [Tools settings](/guide/chat-ui#tools); other MCP clients have their own way to disable a tool).