Separate Personal and Work AI Systems with Minimal Information Handoffs

Design separate personal and work AI entry points, share only the constraint that work needs, and do not mistake Profiles for security isolation.

# Separate Personal and Work AI Systems with Minimal Information Handoffs

One AI system can handle personal tasks and work projects, but being able to access information does not mean that information belongs in the same context. A more careful design uses parallel personal and work entry points. When a personal event affects work, hand over only what changes the work plan—for example, an unavailable time window—not the reason or the full conversation.

Privacy boundary: This article addresses information, role, and state separation. A separate Profile is not operating-system-level isolation and cannot, by itself, prove privacy or security.

Why “can be combined” is different from “should be combined”

Personal context rarely belongs in work planning. A work system does not need a person's complete private background to move a deadline. If every conversation is written into one persistent memory, it becomes difficult to separate delivery-relevant information from private context—and difficult to withdraw information that should no longer be visible to work.

Hermes Profiles can have separate memory, sessions, configuration, credentials, and gateway state. Hermes also distinguishes a Profile from a workspace and a sandbox. Profiles can separate role and state, but they do not automatically provide a filesystem security boundary. [Profiles documentation](https://github.com/NousResearch/hermes-agent/blob/main/website/docs/user-guide/profiles.md)

Parallel entry points with minimal handoffs

The model uses two parallel entry points. A personal assistant handles a personal Board and memory; a work chief of staff handles new work and existing specialists. They do not automatically read each other's memory. If a personal event affects work, the handoff contains only work impact, effective period, minimum constraint, and permission to retain it.

| Field | Example | | --- | --- | | Work impact | No meeting or deadline work in a given period | | Effective period | One specified day, 13:00–17:00 | | Personal reason | Not transferred | | Retention | Store only as a scheduling constraint, not in work memory | | Approval | Transfer only after explicit user consent |

The work system may reschedule Tasks. It should not demand the private reason. If the constraint is canceled, the handoff should expire rather than remain as a permanent personal label.

Three boundaries to design

1. Classify information first

For each input, decide whether it is a personal task, a work task, a temporary conversation, or a cross-boundary matter. Not every conversation deserves durable memory.

2. Limit the use

Personal data serves personal tasks; a work Board holds delivery-relevant information. Hermes configuration can require approval before writing to memory, making it a useful human checkpoint for sensitive retention decisions. [Configuration documentation](https://github.com/NousResearch/hermes-agent/blob/main/website/docs/user-guide/configuration.md)

3. Make the handoff explicit

The person initiating the handoff should state the work impact, duration, minimum constraint, and retention permission. The goal is not to synchronize every memory; it is to allow a necessary scheduling adjustment without exposing unrelated private information.

Do not let role names conceal security limits

If several Profiles share a runtime or mounted directory, this design should be described as role and state separation—not complete privacy isolation. High-sensitivity cases still need a separate assessment of containers, data mounts, OS identities, credentials, and audit records. Hermes Docker guidance also warns against having multiple containers write to the same data directory. [Docker documentation](https://github.com/NousResearch/hermes-agent/blob/main/website/docs/user-guide/docker.md)

Limits and next step

Start with a non-sensitive personal item and observe whether the model reduces unnecessary disclosure, how much friction it creates, and how often people need to correct classifications. Define information that must never be transferred automatically, then use the four-field handoff template. Convenience does not mean the assistant knows more; it means it knows exactly enough when it needs to.