Hermes Agent Community Reports a Gap in Visual History: Gateway Conversations May Lose Tool-Loaded Images

On October 20, a multimodal session report appeared on the Hermes Agent GitHub: in the Discord gateway, native images loaded through vision_analyze may be available to the model only for the current turn. When the next message arrives, the history may retain only text and a [screenshot] placeholder. The reporter observed that the model then retracted an earlier, correct image interpretation. This remains an individual case and does not establish that all gateways or models are affected. Community report

This appears to differ from the documented tool behavior. The documentation says that when the main model and provider support images in tool results, vision_analyze returns pixels directly, avoiding the need for an auxiliary vision model to describe the image. Embedded images should also be sent again with subsequent requests in the conversation. vision.embed_target_bytes controls the ongoing image payload overhead and defaults to 256 KB. Vision feature documentation

The reporter locates the issue at the boundary between persistence and history reconstruction: _durable_content converts multimodal tool results into text summaries, replacing image content with placeholders. When the gateway loads history from session storage on the next turn, it can no longer restore the pixels. More misleadingly, the remaining text still says that the image was loaded and can be viewed directly. The reporter suggests two possible approaches: preserve a reference that allows the image to be attached again, or clearly indicate that the image was viewed previously but its pixels are no longer available. Issue analysis and suggestions

The engineering takeaway is that an agent’s cross-turn reliability with multimodal input also depends on how its history is serialized. Based on the mechanism described above, inconsistency between a model’s answers across turns may result from evidence being lost during storage and reload. The official documentation also describes vision.replay_recent_images, which handles replaying user attachments in history. It discusses this separately from tool results, so it cannot be assumed to solve this issue. History image replay documentation

Next, watch whether any fix covers persistence of tool images, cross-turn reloads, and cache invalidation scenarios, and whether a deployable version is released. Operators can start by asking successive questions about the same image and checking whether later requests still include its image content. A correct answer on the first turn alone is not enough to confirm that the full visual workflow is reliable.