Hermes cron jobs now persist memory between runs

Hermes Agent v0.21.0 adds persistent memory to cron jobs. Scheduled agents now carry state between runs instead of starting fresh each invocation.

Before this release, every cron tick loaded only the immediate prompt and any context supplied at dispatch. If a monitor reported a filesystem change at 2pm, the 3pm run had no record of that report. The agent rediscovered its history from scratch or relied on external logs.

The release introduces four connected behaviors across four merged PRs. PR #91447 lets cron agents load and update persistent memory like interactive agents. PR #80774 adds a continuity flag that carries each run's output into the next invocation. PR #81139 gives every job a durable notepad scratchpad. PR #81138 implements monitor-mode jobs that skip the LLM entirely when the watched signal has not changed.

The monitor-mode skip is the most concrete optimization. When a cron job polls a file, URL, or system metric and detects no change, the runtime bypasses the model call and returns the cached result. That removes inference cost for the no-op case, which is the majority case for most polling loops.

The implementation uses hash-suppressed change detection. Each monitored target gets a hash of its last-seen state. On the next tick, the runtime computes a fresh hash and compares it to the stored value. If they match, the job exits without calling the model. If they differ, the job proceeds normally and updates the stored hash.

The durable notepad uses a key-value scratchpad that survives scheduled runs. This is separate from long-term memory. It stores ephemeral state the job needs between ticks: last-checked timestamp, last-reported hash, counters, or any data the prompt asks the agent to remember. The notepad is scoped to the job, not shared across jobs in the same profile.

The continuity flag changes how cron output is consumed. By default, each cron run is isolated. With continuity enabled, the agent receives the previous run's output as part of its context. This lets a monitor deduplicate alerts: if it already reported a filesystem change, the next run can compare the current state against the previous output and suppress redundant notifications.

Together these changes turn cron jobs from fire-and-forget triggers into stateful agents with session continuity. A nightly backup checker can remember which files changed last night. A system monitor can suppress repeated alerts for the same condition. A data pipeline can carry a cursor or watermark between runs without external storage.

This release also includes other operational improvements. Subagent delegate_task gained live orchestration: list running children, steer one mid-flight, or stop it early. Defaults raised to 250 iterations and 10 concurrent children. The MCP server management surface became a desktop dashboard with health checks and fleet cost overlays.

The Pantheon Release rolls up changes from v0.20.1 through v0.20.6. The release notes document approximately 5,800 commits, 2,475 merged PRs, and 2,100 closed issues across 760+ contributors since v0.20.0.

[^1]: NousResearch. "Hermes Agent v0.21.0 (v2026.8.31)." GitHub releases. August 31, 2026. [^2]: @NousResearch. "Hermes Agent v0.21.0: The Pantheon Release." Thread on X. August 31, 2026. Unroll. [^3]: teknium1. "feat(cron): cron jobs now load and update persistent memory." PR #91447. Merged August 21, 2026. [^4]: teknium1. "feat(cron): continuity=true." PR #80774. [^5]: teknium1. "feat(cron): monitor-mode jobs." PR #81138. Merged August 2026. [^6]: teknium1. "feat(cron): per-job durable notepad." PR #81139.