Memory Card

Memory card is the atomic unit of any agent memory system that deserves user trust. It holds one single fact preference or observed pattern along with the metadata that makes it actionable and safe. That metadata includes the creation timestamp the scope that limits where the card applies the source that shows whether the user explicitly created it or the agent inferred it from behavior and the expiration policy that prevents stale data from poisoning future interactions. Cards are shown to users as discrete interface objects complete with scope chips decay timers and one click edit delete and export buttons. This turns memory from an invisible backend process into a collaborative surface that users can curate like they curate their notes or files. The four memory types map to cards in different ways. Preference cards tend to be explicit and long lived. User fact cards accumulate quickly and require strong visibility because they feel personal. Work in progress cards need aggressive expiration tied to project milestones. Behavior signal cards are the trickiest because they come from inference and must carry clear signals that they are not user stated facts. Every card follows the five trust principles or it becomes a liability. Visible means the card appears in the memory inspector and gets cited with a link in responses. Editable means the user can rewrite the text not just delete it. Scoped prevents cross contamination between work and personal contexts. Expirable stops the slow accumulation of junk. Exportable lets the user take their cards to another product in clean markdown or JSON that reads like notes not code.

A memory card is not a database record hidden from view. If users cannot find it in one click it fails the first trust principle and creates the surprise problem that plagued ChatGPT after its 2025 memory launch. It is not an entire conversation log or vector store dump. Those approaches ignore the curation problem and leave users with no way to manage what the agent knows. A memory card is not a permanent fixture unless the user declares it so. The creep failure mode happens when cards accumulate without caps or decay timers leaving users with 400 entries after three months of use and no easy way to clean house. It is not a silent inference that the agent acts on without disclosure. Behavior signals must be presented as cards for approval before they shape responses or users feel watched. It is not a replacement for good onboarding. Cards cannot fix a product that never asked the user their basic preferences in the first place. Finally it is not a lock in tactic. Products that make cards hard to export create the exact resentment that drives users to competitors with better memory portability like the .cursorrules approach in Cursor.

Take the 2026 experience of a freelance brand strategist using Claude for client work. She creates a project for a wellness brand and tells the agent our voice is empathetic but never condescending and we always reference evidence from peer reviewed studies. The agent generates a preference card with those exact words after a confirmation toast that explains the scope is limited to this project. Later when she pastes research notes the agent creates work in progress cards for key findings with decay timers set to 30 days after project close. In the memory inspector she sees all cards laid out like voxel tiles with filters for type and source. One inferred behavior signal card says this user prefers shorter replies after 5pm based on her reply patterns. She clicks the why did you decide that button and sees the audit trail of three previous conversations where she asked for brevity in evening sessions. She edits the card to make it a hard preference rather than time based and adds it to her cross project favorites which is a newer Claude feature that lets certain cards roam beyond their original scope. When she switches to a different client project for a tech startup the wellness cards stay behind thanks to strong scoping. This stands in stark contrast to early ChatGPT memory that created cards from every conversation including a one off joke about her coffee addiction that resurfaced months later in a work context causing the surprise that made her switch products. The freelancer exports her cards monthly as a markdown file titled my agent knowledge base which she imports into new tools without friction. The same pattern appears in Cursor where a developer adds a rule to .cursorrules for always using TypeScript strict mode and the card appears in the sidebar with a visible git history showing exactly when it was added and by whom.

Implement memory cards when your product expects repeated use and the value grows with shared history. They shine in creative tools where style preferences matter like a designer setting font pairing rules once and having them respected across every new file in a 2026 version of Figma AI. They solve the repetition problem in meeting tools like Granola by creating notebook specific cards that remember recurring attendees and their roles without being re explained. Use them when you can commit to building the supporting features from the design workshop including the inspector screen with bulk actions and the weekly digest that shows new cards created. Memory cards become the moat once models are commoditized because users stay where the agent actually knows them without the memory hole or the creep. Avoid them in high sensitivity domains like mental health agents where the default should be no persistent memory unless the user explicitly builds a card and even then with extra encryption and delete guarantees. Do not use memory cards if your disclosure system is weak because the first time an inferred card surfaces without context you trigger the watched feeling that no amount of later fixes can fully repair. Skip the card approach for products aimed at users who value speed over personalization like quick query tools where adding memory overhead slows the core experience. The teams that launched without addressing the four failure modes in 2025 spent all of 2026 playing catch up on trust.

The best memory card systems make the user the author of their own agent instead of the subject of its observations.