Hermes Agent v0.21.0 update gives cron jobs memory
Released on August 31, 2026, Hermes Agent v0.21.0 introduces the continuity feature, allowing scheduled cron jobs to retain context and reduce API costs.
Stock photo for illustration only, not from the actual event
- Hermes Agent v0.21.0 adds the continuity feature for scheduled cron jobs to retain context.
- Reduces LLM invocation costs by running initial check scripts before calling AI models.
- Introduces no_agent mode for automated tasks that operate entirely without model calls.
- Supports routing cron results directly into chat rooms and chaining multiple tasks together.
A classic issue with automated cron jobs is their lack of memory. Tasks scheduled to run every morning to check data typically start from a blank slate every single time, resulting in repetitive reports that require users to manually read and filter through them. The release notes for Hermes Agent describe these recurring scheduled tasks as having the memory span of a goldfish.
To tackle this limitation, the development team released version v0.21.0 on August 31, 2026, bringing a vital feature that allows scheduled tasks to load persistent memory and update it much like an interactive chat agent. This upgrade ensures that jobs no longer begin from scratch, allowing them to review past logs and overwrite or append information accurately.
The core mechanism driving this change is the continuity variable, which feeds the output of the previous cycle directly into the next. For instance, consider a monitoring task that checks a competitor's webpage every hour. Without memory, the task reports everything it sees, which rarely changes. With persistent memory enabled, the report shrinks to highlight only the lines that have actually shifted. Each task maintains its own dedicated scratchpad, separated cleanly from the main memory of the system.
Separating the persistent scratchpad from core memory and introducing a preliminary check before invoking the LLM represents an exceptional architectural choice for cost management. In typical AI systems, querying models during every execution cycle regardless of new data generates massive overhead. Filtering data via scripts beforehand effectively eliminates unnecessary API expenses.
Stock photo for illustration only, not from the actual event
The workflow is optimized to minimize unnecessary resource consumption. Before making a request to the LLM, the task executes a checking script. If the script determines that nothing has changed, the cycle concludes immediately without triggering the LLM at all, greatly reducing costs for frequent checks that rarely encounter events. Furthermore, the release features a no_agent mode for tasks where the script handles everything, such as checking if RAM exceeds 85% and sending a Telegram alert without involving any AI model.
On the security and configuration front, the system incorporates readiness checks prior to creating any task, verifying provider keys, attached skills, and destination endpoints. If any check fails, the task flags a configuration error, sends a single alert, and avoids burning quota silently in the background. The subsequent v0.21.5 release, rolled out on September 24, 2026, further refined the overall platform stability. Users managing scheduled routines are encouraged to verify their continuity settings to transform fire-and-forget scripts into context-aware automated workflows.
Source: Dev.to
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment