Scope Chip

What it is. The scope chip is the small colored pill that tells you and the agent exactly which memory bucket is active. This conversation. This project. This workspace. All my work. Everything. It lives on every memory card in the inspector, rides along on every agent reply that cites stored context, and appears as a selector whenever new memory is created or a fresh session begins. The chip makes the Scoped trust principle concrete instead of theoretical. Tap it and you change the boundary on the fly. Filter the memory inspector by it and every related card lines up with its decay timer and audit trail attached. In 2026 products that got this right turned memory from a liability into a moat. Products that ignored it watched users flee after one bad cross contamination incident. The chip pairs with memory cards as the atomic unit of trustworthy design. It uses color coding, one click editing, and clear defaults so users never have to ask how does it know that. Designers who treat the chip as infrastructure rather than decoration ship agents people actually trust with their work their preferences and the small embarrassing facts that make the agent feel personal.

What it isnt. A scope chip is not a decorative tag slapped on for visual interest. It is not a tooltip that appears only on hover. It is not a backend flag that users discover after the fact through a disappearing toast. It is not the same as a folder from 2010 software because folders do not travel with individual facts into agent replies or audit trails. It is not a global on off toggle and it is not something the system infers then hides behind doublespeak about improving the model. If changing the chip requires three menus or a support ticket it has already failed. It is not a crutch for sloppy defaults. Dumping everything into an Everything bucket and then adding chips later just creates an obvious mess that users screenshot and mock. The chip must carry real behavioral weight or it becomes theater that erodes trust faster than no chip at all. Teams that shipped memory features in 2025 without visible editable scope chips learned this lesson in public very quickly.

Concrete example. Claude projects in late 2025 gave the clearest proof. You create a project called Q3 Brand Refresh. Every uploaded brief every custom instruction and every learned preference inside it carries a green scope chip that reads This project. Open the memory inspector and cards group cleanly under their chips. When the agent cites your tone guidelines in a response the chip appears inline next to the citation so you see This project and never wonder why that detail surfaced. Switch to your separate Therapy Journal project and the chip turns purple. The two never mix. No brand voice leaking into private entries. Users consistently rated this higher than any other memory system that year because the boundary felt user drawn not system imposed.

ChatGPT launched memory in February 2025 with zero scope chips and the backlash was immediate. The model would silently store a preference for bullet points from a casual email thread then apply it months later while you drafted a sensitive performance review. The only signal was a tiny Memory updated toast that vanished in two seconds. One product designer discovered his AI referencing a joke about his CEO from a throwaway chat while prepping a client pitch deck. Screenshots spread across LinkedIn. The memory drawer showed hundreds of entries with no visible scope no easy way to filter and no way to promote or demote them. OpenAI patched it in 2026 but trust recovery took quarters. Cursor took the opposite path. Its .cursorrules file lived inside the repo and acted as one giant scope chip for This project by definition. When Cursor added a visual memory inspector in version 0.42 every rule became a card with its own editable chip. Developers could promote a TypeScript preference from This project to All my work in one tap. A design tool called Context Keeper copied the pattern for Figma in 2026. Each component could hold memory cards about stakeholder feedback. The default chip read This file but could expand to This team workspace. The attached audit trail showed exactly which chips influenced each suggestion. Trust scores jumped 40 percent in beta tests because users finally saw the guardrails instead of guessing at them.

When to use when not to. Use scope chips in any product that supports multiple concurrent contexts or mixes the four memory types. Put them on every memory card. Surface them in agent replies that pull context. Require explicit chip selection for the first ten automatic saves so users own the decision before The Surprise kicks in. They become essential the moment you move beyond single chat interfaces. Combine them with decay timers so a This conversation chip auto expires when the project ships. They fight The Creep by letting users bulk delete everything under one chip. They fight The Lock In by making exports filterable by scope before users leave for another tool. Run the six step workshop and spend real time defending the default chip for each memory type. If the defense sounds weak change it before any code ships. The memory inspector without scope chips is a refrigerator without shelves. Everything piles up and you waste hours digging.

Do not use them in single purpose tools where the container itself defines the boundary. Granola notebooks need no chips because the notebook is the scope. A dedicated personal journal app with one user and one context can skip them too. Avoid them in consumer mobile products where every extra UI element hurts adoption and the use case stays narrow. Never treat chips as cosmetic polish added after launch. They must be functional from day one or they create false confidence in broken boundaries. The teams that win the next three years will treat scope chips like they treat search or onboarding with dedicated reviews metrics and iteration instead of bolting them on as an afterthought.

Scope chips turn every memory decision from an act of faith into an act of informed consent.