Decay Timer

A decay timer is the visible label on every memory card that declares exactly when that memory loses its influence over the agent. It turns the expirable principle into pixels users can see edit and argue with instead of discovering six months later that a sarcastic comment from February is still shaping their project brief responses in July. The label might read expires in 18 days or drops when this sprint ends on March 12. Clicking it opens a quick editor with presets tied to real events plus a natural language field where users type until brand refresh or when behavior shifts twice. The system parses it links it to the calendar or usage patterns and updates the agent behavior the moment the timer hits zero. This is not decoration. It is the feature that stops memory from becoming a digital hoarder that keeps every receipt and slowly poisons every future reply. Designers own this because engineers default to invisible relevance scoring that no user can inspect or control. Your job is to force these timers into the memory inspector make them as prominent as the fact itself and tie them to the four memory types so preferences stay sticky while work in progress dies on schedule.

A decay timer is not backend TTL logic that runs silently and surfaces only in error logs. It is not a delete button with extra steps. It is not the vague archive folder most products slap on when they realize users hate infinite storage. It is not optional. It is not the same slider applied to every memory type and it is definitely not something you explain in a tooltip that says our model handles freshness. Teams that hide expiration behind we use ML to decide create the exact surprise and creep problems the 2025 products became infamous for. Real decay timers are legible editable and tied to observable behavior change. If the timer expires but the agent still references the memory you have built theater not memory design.

ChatGPT proved the point in early 2025 when its memory feature shipped without decay timers. Users casually mentioned a bad breakup in one late night thread only to watch the model awkwardly avoid any relationship adjacent language in work copy for the next eight months. The March 2026 update finally added decay timers as clock badges on every card in the memory inspector. One card read user hates pineapple on pizza with a 30 day timer that users could extend to permanent or shorten to expires after this grocery list. OpenAI also shipped weekly digests showing every new timer created that week. Complaints about stale memory dropped sharply. Claude took the opposite path in projects. By late 2026 every uploaded file or instruction set required a decay timer at creation time with options like this campaign only or until Q3 results ship. The project dashboard showed a fading timeline of all context with visual indicators that turned from green to amber as timers wound down. A design team at Figma itself used this to manage client brand work where old visual direction cards auto expired the day after new guidelines launched so the agent stopped suggesting outdated aesthetics. Cursor pushed it into code. Their .cursorrules files now support comments like always open with questions // decay-timer: 45d or on 3 rejections. The Cursor memory pane renders these as editable visual timers that update the source file on change. Granola applied the same thinking to meeting notebooks where every extracted action item ships with a default 21 day timer because last quarters priorities become noise in this quarter. The hourglass icon next to each insight fills gradually and taps straight into the edit timer flow. These four concrete implementations across chat code notebook and project interfaces show that decay timers must adapt to the product but never disappear from view.

Use decay timers on behavior signals and work in progress context where relevance has a natural half life. Put them front and center in the memory inspector so users can sort by soonest to expire and kill the creep before it starts. They become essential in any agent that spans work and personal contexts or runs for longer than one quarter. Combine them with scope chips and audit trails so every agent reply that uses memory also shows the remaining timer life. They are mandatory in the six step memory workshop during the expiration policy step. Set smart defaults per memory type then let users override without friction. Do not use them on core user facts that form identity or on explicit permanent preferences like accessibility settings and tone defaults. Skip them entirely in one shot tools where nothing should persist past the session. Never treat them as a fix for bad scoping. A therapy bot that remembers work deadlines even with a 7 day timer still feels like it is spying. Overcomplicating the timer UI with cron syntax or twelve presets will make non technical users run for the off switch.

Decay timers are how you prove your agent can forget on purpose instead of pretending it never happened.