How BenAgent Keeps Hermes and OpenClaw Context Stable

Conversation continuity is what makes a desktop agent useful after the first prompt. BenAgent owns that continuity locally: completed user and assistant turns are persisted in SwiftData, bounded, ordered from oldest to newest, and embedded in the next request.

Hermes requests

For Hermes Chat Completions, BenAgent sends one system message and one user message. For Hermes Responses, it sends system instructions and one user input item. Both carry the same local-history envelope; BenAgent does not send X-Hermes-Session-Id or previous_response_id on later turns.

OpenClaw requests

For OpenClaw Chat Completions, BenAgent sends one system message and one user message. It does not send a conversation-derived user value or reuse a gateway session key. Follow-up context comes from the history inside the user-message envelope.

What breaks context

• Changing provider settings while a reply is still in flight.
• Starting a new local conversation before the current request returns.
• Expecting failed or cancelled turns to appear in completed history.
• Expecting the raw bytes of an old image to be retransmitted instead of its history reference.

Good session habits

• Let a request finish before switching providers.
• Use Reset Session intentionally when you want a new local conversation.
• Keep separate test prompts for Hermes and OpenClaw so endpoint problems do not look like memory problems.
• When debugging, inspect the one user-message envelope and verify its history is ordered oldest to newest.

Test session continuity

Remember this test phrase for the next message: blue local bridge.

What test phrase did I ask you to remember?