Portable AI Agent Memory: What Should Move When Users Switch Agents?

When transitioning from one AI assistant to another, users often encounter the challenge of reconstructing their preferences, projects, and working relationships from scratch. This can be seen as a loss of continuity and efficiency. The issue at hand is whether AI agents should be able to transfer their accumulated context when users switch applications.

AI agents maintain various types of information which have different levels of portability. Active conversation and task state, durable facts and project knowledge, user preferences, historical decisions, and authorization-related records are all stored within an agent. Some of this information, like a user's preference for concise technical reports, remains useful when switching agents.

However, other data, such as active execution checkpoints or historical approvals, may not be directly transferable due to their dependence on specific runtimes, configurations, or permissions.

To address this problem, there are two key design principles proposed in the source material. The first principle suggests preserving useful context during memory transfer without automatically granting operational authority to the new agent. For instance, if an agent remembers that a manager approved a purchase order, this historical fact should not automatically grant the new agent the permission to execute similar actions.

The second principle proposes separating memory content from memory authority. This means distinguishing between the actual data stored in the agent and the permissions associated with that data.

A minimal portable memory schema is proposed to facilitate practical memory transfers. This schema includes fields for schema version, memory ID, owner ID, memory type, content, source, creation date, validity period, and export class. The export class indicates whether the memory record should be considered for export, reviewed before export, or excluded from the migration process.

To implement the memory transfer process, Python code is provided to classify memory records based on their export eligibility. This involves checking for valid expiration dates, determining the memory type (such as preference or authorization), and classifying the record as eligible for export, requiring review, or being excluded from the migration process.

By separating memory content from memory authority and implementing a structured export policy, users can have more control over their AI agent's memory during application transitions. This approach ensures that useful contextual information is preserved while minimizing the risk of transferring unauthorized permissions or outdated data.