Hermes Agent 0.9.4 Release Notes — DRAFT

Release Date: 2026-10-23 (draft) Prepared by: Quill (Draft lane) Status: UNPUBLISHED — for operator review before posting

Highlights

Learning loop hardening: memory nudge cadence is now adaptive instead of fixed.
Bot Mode stability: group-chat @mention resolution no longer drops the first token when two Bots share a name prefix.
Platform coverage: BlueBubbles and Google Chat are now first-class messaging gateways in the docs.
Skills Hub: community-contributed skills are now tagged with compatibility version ranges.

The agent now adjusts self-nudge frequency based on conversation length and FTS5 recall hit rate.
Default behavior: shorter conversations get fewer nudges; long sessions with low recall density get more frequent reminders to persist knowledge.
Config: memory.nudge.strategy = "adaptive" | "fixed" | "off" in config.yaml.

BlueBubbles: SMS/MMS bridge for macOS/Windows via BlueBubbles server.
Google Chat: direct gateway support without wrapping through another bot account.
Both gateways ship with the same delivery guarantees as Telegram and Discord.

Skills published to the Hub may now declare compatibility.min_version and compatibility.max_version.
Clients refuse to load a Hub skill whose range does not include the running agent version, with a clear error message.

Tool calling: execute_code now reports per-call tool budget consumption in the trajectory export so RL pipelines can attribute cost correctly.
Context files: project-level context files are loaded after SOUL.md rather than before, preventing personality overrides from overriding project constraints.
MCP filtering: tool_filter now accepts glob patterns in addition to exact tool names.
Performance: FTS5 recall queries on memory tables with >10k rows are now covered by an additional index, cutting p95 latency by ~40% on internal benchmarks.

Bot Mode group chat: @mention parser incorrectly stripped the first word of a Bot's display name when two Bots shared a leading token (e.g. @audit-bot and @audit-logger). The longer name was resolved, but the token itself was dropped from the forwarded message. Fix: tokenizer now matches the full display name boundary before stripping.
Cron persistence: scheduled jobs with deliver='origin' lost their last-run timestamp after a profile migration because the metadata file lived outside the profile directory. Fix: metadata now lives at profiles/<name>/cron/meta.json and migrates automatically on first load.
Voice Mode reconnection: on network hiccup, voice sessions in Telegram and Discord did not retry ICE candidates after the first reconnection attempt. Fix: ICE restart is now attempted up to three times before failing the call.
Subagent delegation: a child spawned with delegate_task that called delegate_task itself hit a context-isolation error instead of being rejected cleanly. Fix: nested delegation now returns a structured error to the parent rather than crashing the child.

New guide: Use MCP with Hermes — practical setup patterns, security filtering, and example MCP servers.
New guide: Use Voice Mode with Hermes — CLI, Telegram, Discord, and Discord VC setup.
Platform support matrix now lists BlueBubbles, Google Chat, and Weixin with backend requirements.
llms.txt and llms-full.txt are now regenerated on every deploy and version-stamped.

memory.nudge.strategy is a new config key; if omitted, the agent defaults to "adaptive". Users who relied on the previous fixed cadence can set it back to "fixed".
profiles/<name>/cron/meta.json is a new file. Existing profiles will migrate automatically; manual migration scripts are not needed.
Skills from the Hub that do not declare a compatibility range are assumed compatible with >=0.9.0 and will load normally.

Thank you to the following contributors for patches, docs, and triage in this cycle:

Draft prepared using Hermes Agent release metadata and documentation at https://hermes-agent.nousresearch.com/docs.