GitHub Repo
Hermes Agent v0.22.6 Released with Background Compression Summaries and Pre-Inference Blocking
Hermes Agent released v0.22.6 on November 2, bringing background preparation of conversation summaries and a pre-inference blocking hook into the stable release. The former must be enabled explicitly and may increase summary costs; the latter blocks at the level of each user turn.
Nous Research released Hermes Agent v0.22.6 on November 2, bringing together about 776 merged PRs since v0.22.5. Two changes stand out for developers of agent systems: optional background preparation of compression summaries, and the pre_llm_call hook, which can block a turn before inference. A full, consolidated changelog is expected with v0.23.0. Release announcement
Background summaries are controlled by compression.prepare_ahead, which is disabled by default. As a conversation approaches the compression threshold, the system can generate a summary during tool execution or between turns. When automatic compression is triggered, it uses the prepared summary if it still corresponds to the same set of messages, while keeping newer messages in the uncompressed tail. If the background task is not finished, the system waits within the existing compression timeout to avoid generating the summary again. If the summary is no longer valid, it falls back to the original process. Configuration documentation
Based on how this mechanism works, it could move summary latency off the user's critical waiting path, but it incurs an auxiliary model call that may go unused. It also does not apply to manual /compress, plugin context engines, or native/app-server compression. If a memory provider appends content to summaries, the feature disables itself after the first compression. Operators should compare wait times with summary usage rather than assume that enabling it will reduce costs. Feature limitations
Another change lets plugins return a blocking action that ends a turn before the main tool loop begins. The documentation says the hook runs after context compression, once per user turn—not once per model API call. When a block is applied, the specified message replaces the model response, no model call is made for that turn, and the result includes the reason for the block. This makes the hook suitable for checks at the start of a turn; engineers still need to handle controls for subsequent tool steps separately. Hook documentation
When upgrading, check the distribution channel: run hermes update for the CLI; update the Microsoft Store version through the Store; and use the corresponding images for Docker and Hermes Cloud. The desktop version installed through an installer is not updated with this release and requires the next bundled release. These features are included in a stable release, but no measurements are available to support claims about the extent of any latency improvement. Upgrade instructions