Memory Inspector

The memory inspector is the dedicated full screen dashboard that shows users exactly what their agent has stored about them. It organizes memory cards by scope with tabs for global project specific and conversation specific items. Filters across the top isolate the four types so a designer can view only behavior signals or only user facts. Each card lists the exact text the agent uses a timestamp the source conversation link a scope chip a decay timer and one click buttons to edit delete pin or export. The inspector links directly to audit trails so every agent reply carries a button that opens the relevant cards. This screen forces concrete decisions on visibility editability scoping expiration and exportability. It turns the black box of agent memory into a tangible document users coauthor and control. Products that shipped this in 2026 treated it as the primary surface for power users instead of a buried settings page.

The memory inspector is not the tiny drawer ChatGPT launched in early 2025 that showed five entries then vanished behind a disappearing toast. It is not chat history. It is not a read only log of system inferences. It is not the project knowledge panel in Claude that ignores behavior signals. It is not a plain .cursorrules text file even though that pattern scores high on legibility. It is not a backend debug view or a JSON dump. It is not an afterthought modal added after users scream on Twitter.

A concrete example comes from Palette the AI design tool that launched its inspector in August 2026. The team ran the six step workshop before writing any backend code. The resulting screen defaults to a kanban layout with columns for the four memory types. A preferences card reads User always wants three rounds of revision with a scope chip saying Global and a permanent timer. A behavior signals card states User rejects first two logo concepts but engages on iteration four pulled from seven Figma sessions in Q2 2026 with links back to each. One designer found a user facts card wrongly listing their boss as Steve instead of Stephanie edited it in plain text and watched the agent correct course the next day. The export produced a markdown file titled Palette Memory Export that read like clear notes instead of a database dump with headings for each scope and bullet lists users could drop into Claude or Cursor. Another example is the 2026 upgrade to Memoir the AI writing companion. Its inspector added a heatmap showing which memories the agent actually referenced in the past 30 days. A writer used it to expire a work in progress card tied to an abandoned novel so it stopped contaminating new projects. These screens fixed the creep with bulk delete the surprise with source links and the memory hole with explicit pinning.

Ship the memory inspector before you ship any persistent memory. Use it when your agent stores anything across sessions because it prevents the four failure modes and satisfies all five trust principles in one interface. Run every design review against the inspector to catch stale data or scope leaks. Check it weekly like you check analytics. Do not use it for stateless tools like one shot image generators or temporary calculators that forget everything at session end. Never hide it behind paywalls or nested menus and never launch memory features without it or you will spend the next year issuing apologies instead of updates.

Build the memory inspector first and your users will trust the agent with the messy truths they hide from everyone else.