Showcase: hermes-companion — persona evolution + emotion state + ambient processing + proactive check-ins as a single zero-patch plugin (refs #11919 #22136 #9645 #13529)

[gejifeng]

TL;DR

I built hermes-companion, a zero-patch plugin that bundles several long-standing community requests (persona evolution, ambient processing, proactive check-ins, emotional state exposure) into one cohesive "AI companion" layer on top of Hermes v0.12+.

Repo: https://github.com/gejifeng/Hermes_Soul_patch

This is a showcase / reference implementation, not a feature request — opening this so people watching the related issues below can find a working example, and so the Hermes team / community can give feedback or pull anything useful upstream.

---

What it implements (and which existing issues it relates to)

Feature	Relates to
EMOTION_STATE.md injected into system prompt; valence/arousal/dominance auto-decays each turn; auto-inferred from conversation via the auxiliary LLM; drops on tool failure; /mood, /mood-set commands	#13529 (Agent Activity API & Emotional State Exposure)
daily_seed (method C): fixed anchor events + persona-driven LLM augmentation written into world_state/events.json at each day rollover — SOUL.md effectively evolves day-to-day through lived events instead of being static	#11919 (SOUL.md should evolve with usage)
Background heartbeat thread: idle-time emotion drift, periodic daydream candidates, arousal-threshold triggered proactive messages	#22136 (Ambient "Daydreaming"), #9645 (Optional, Configurable Proactive Check-Ins)
LLM-callable agenda tools (agenda_add, agenda_done, agenda_ambient), /recall for past-day brief	#12323-style daily summary use cases
Time awareness sub-module delegates entirely to hermes_time._resolve_timezone_name() — respects HERMES_TIMEZONE env / ~/.hermes/config.yaml — no duplicated logic	sister extension https://github.com/gejifeng/hermes-time_perception-extension
All wired through the official plugin hook surface (pre_llm_call / post_llm_call / post_tool_call / on_start / on_shutdown) + slash commands — zero patches to Hermes core	—

87 unit tests, no external deps in the test path. Verified against the v0.12.x plugin hook surface.

---

Why post here

The features above each have their own open issue/discussion, but nobody (that I could find) has tried to wire them together end-to-end as a single plugin. Maybe this saves someone time, or surfaces gaps in the current plugin API that would be worth addressing upstream — e.g.:

• Plugin-side proactive injection currently has to fall back to either ctx.inject_message from inside the process or a file-queue picked up by an external watcher; a first-class "plugin → gateway outbound" path (similar to Plugin system lacks webhook delivery extension point — plugins must patch core source code #15006) would make this much cleaner.
• post_llm_call is the only practical hook for cheap auxiliary-LLM emotion inference; if auxiliary_client shape changes between versions, plugins like this break silently. A small stability contract there would help.

Happy to split anything useful into separate, focused issues if there's interest.

---

中文版

我做了一个零侵入的 Hermes 插件 hermes-companion，把社区里几个长期 open 的 feature request 拼接成一个完整的"AI 陪伴层"，跑在 Hermes v0.12+ 上。

仓库：https://github.com/gejifeng/Hermes_Soul_patch

这不是新的 feature request，而是一个 showcase / 参考实现，发在这里是让关心下面这些 issue 的人能找到一个可运行的样例，也欢迎 Hermes 团队和社区给反馈或抽取有用的部分到上游。

对应到现有 issue：

• [Feature]: Agent Activity API & Emotional State Exposure #13529 — 注入 EMOTION_STATE.md 到 system prompt；valence/arousal/dominance 每轮自动衰减；通过 auxiliary LLM 在 post_llm_call 中自动推断；工具失败时 valence 下降；提供 /mood /mood-set。
• SOUL.md should evolve with usage, not just initialize once #11919 — daily_seed（方法 C）：固定锚点事件 + 由 SOUL 驱动的 LLM 增强写入 world_state/events.json，让 SOUL 通过"日常经历"逐日演化，而不是一次性初始化。
• Feature Request: Ambient "Daydreaming" (Idle-State Background Processing) #22136 / [Feature] Optional, Configurable Proactive Check-Ins #9645 — 后台 heartbeat 线程：idle 期情绪漂移、周期性 daydream 候选、arousal 阈值触发主动消息。
• 与时区扩展兼容：完全委托 hermes_time._resolve_timezone_name()，不重复造时区逻辑（参见姐妹仓库 https://github.com/gejifeng/hermes-time_perception-extension）。
• 全部通过官方插件 hook（pre_llm_call / post_llm_call / post_tool_call / on_start / on_shutdown）+ slash command 实现，Hermes 核心零修改。

87 个单测，测试路径无外部依赖。已对 v0.12.x 插件 hook 表面验证。

做这个过程中发现两个可能值得上游处理的小痛点：

1. 插件侧发起主动消息目前只能 ctx.inject_message 或落文件队列+外部进程消费，缺一条"插件 → gateway 出站"的官方通路（与 Plugin system lacks webhook delivery extension point — plugins must patch core source code #15006 思路类似）。
2. post_llm_call 是做廉价 auxiliary LLM 情绪推断的唯一可用 hook；如果 auxiliary_client 形状跨版本变动，这类插件会静默失效。有个轻量的稳定性约定会很有帮助。

如果其中任何一点有兴趣继续推进，我可以拆成单独的 focused issue。

[TKG-svg]

I'm extremely interested in this—would it be possible for me to use it, if that's alright with you?

[gejifeng]

I'm extremely interested in this—would it be possible for me to use it, if that's alright with you?

I am now using it and feels very good with honcho memory system. You may have a try. If you encounter any problems, I would be glad to communicate with you.

[TKG-svg]

Really?! Thank you so much!

[UniGood]

Cool project — I built something with overlapping goals, hermes-active (github.com/UniGood/hermes-active), running in production for ~2 months now. Some interesting differences from your companion:

• 3D emotion model (valence / arousal / social_need) instead of 2D, with VA decay between beats + LLM re-assessment + dynamic merge every heartbeat
• Multi-factor decision matrix before any send: score = emotion_intensity × time_fitness × silence_factor × frequency_limit, with hard SKIP fallback if the score doesn't clear threshold. The LLM is asked whether to contact, but the matrix decides whether the message actually leaves — so a misbehaving prompt can't bypass the gate
• Heartbeat audit UI — every beat persists full reasoning, context bundle, recall hits, LLM prompt+response, decision score, and post-send emotion state to SQLite. Closing most of [Feature]: Agent Activity API & Emotional State Exposure #13529 with the data already there, just needs an endpoint
• Plain companion UI with mobile layout + animated VA bars + heatmap longing, designed to be glanceable

Setup is a FastAPI + Vue 3 web dashboard (port 18720), reads from state.db only, writes to its own active.db. Heartbeat interval is 5min default, ~30s/beat, ~5KB detail log per beat. Runs well on a 1GB VPS.

Worth comparing: how do you wire the auxiliary_client call for emotion evaluation — same call vs separate from thought-gen? On my side I split them (cheaper/faster eval LLM, slower thought-gen LLM) and it cut p95 from 35s to 22s. Happy to dig into that, or any of the other related issues (#15400 #9645 #13529 #11919) if useful.

[UniGood]

Cool project — I built something with overlapping goals, hermes-active (github.com/UniGood/hermes-active), running in production for ~2 months now. Some interesting differences from your companion:

• 3D emotion model (valence / arousal / social_need) instead of 2D, with VA decay between beats + LLM re-assessment + dynamic merge every heartbeat
• Multi-factor decision matrix before any send: score = emotion_intensity × time_fitness × silence_factor × frequency_limit, with hard SKIP fallback if the score doesn't clear threshold. The LLM is asked whether to contact, but the matrix decides whether the message actually leaves — so a misbehaving prompt can't bypass the gate
• Heartbeat audit UI — every beat persists full reasoning, context bundle, recall hits, LLM prompt+response, decision score, and post-send emotion state to SQLite. Closing most of [Feature]: Agent Activity API & Emotional State Exposure #13529 with the data already there, just needs an endpoint
• Plain companion UI with mobile layout + animated VA bars + heatmap longing, designed to be glanceable

Setup is a FastAPI + Vue 3 web dashboard (port 18720), reads from state.db only, writes to its own active.db. Heartbeat interval is 5min default, ~30s/beat, ~5KB detail log per beat. Runs well on a 1GB VPS.

Worth comparing: how do you wire the auxiliary_client call for emotion evaluation — same call vs separate from thought-gen? On my side I split them (cheaper/faster eval LLM, slower thought-gen LLM) and it cut p95 from 35s to 22s. Happy to dig into that, or any of the other related issues (#15400 #9645 #13529 #11919) if useful.

[UniGood]

Cool project — I built something with overlapping goals, hermes-active (github.com/UniGood/hermes-active), running in production for ~2 months now. Some interesting differences from your companion:

• 3D emotion model (valence / arousal / social_need) instead of 2D, with VA decay between beats + LLM re-assessment + dynamic merge every heartbeat
• Multi-factor decision matrix before any send: score = emotion_intensity × time_fitness × silence_factor × frequency_limit, with hard SKIP fallback if the score doesn't clear threshold. The LLM is asked whether to contact, but the matrix decides whether the message actually leaves — so a misbehaving prompt can't bypass the gate
• Heartbeat audit UI — every beat persists full reasoning, context bundle, recall hits, LLM prompt+response, decision score, and post-send emotion state to SQLite. Closing most of [Feature]: Agent Activity API & Emotional State Exposure #13529 with the data already there, just needs an endpoint
• Plain companion UI with mobile layout + animated VA bars + heatmap longing, designed to be glanceable

Setup is a FastAPI + Vue 3 web dashboard (port 18720), reads from state.db only, writes to its own active.db. Heartbeat interval is 5min default, ~30s/beat, ~5KB detail log per beat. Runs well on a 1GB VPS.

Worth comparing: how do you wire the auxiliary_client call for emotion evaluation — same call vs separate from thought-gen? On my side I split them (cheaper/faster eval LLM, slower thought-gen LLM) and it cut p95 from 35s to 22s. Happy to dig into that, or any of the other related issues (#15400 #9645 #13529 #11919) if useful.