OPEN LINK ↗
// 63d agoOPENSOURCE RELEASE
Hermes Agent self-improving pitch draws skepticism
The Reddit thread pushes back on Nous Research's Hermes Agent pitch, arguing that its memory and skill-building look a lot like the same manual "remember this / save this workflow" loop users already run in Gemini CLI or opencode. What feels different is the packaging: bounded persistent memory, reusable skills, and a deeper user-modeling layer rather than magic autonomous learning.
// ANALYSIS
Hermes isn't reinventing agent memory; it's systematizing it.
- –Hermes stores curated state in `~/.hermes/memories/` across `MEMORY.md` and `USER.md`, so persistence is real even if the learning is narrow.
- –Skills are procedural memory: the agent can turn a repeated workflow into reusable `SKILL.md`, which is useful but still explicitly captured.
- –The self-improving story gets stronger only if Honcho is enabled; by default, the system is more "remember and reuse" than "autonomously improve."
- –Compared with Gemini CLI or opencode, Hermes stands out more as a full agent runtime: CLI, messaging gateways, cron/automation, browser control, and local-first storage.
- –That's the actual moat: fewer repeated instructions, more reusable automation, not magical intelligence gains.
// TAGS
hermes-agentagentcliautomationopen-sourceself-hosted
DISCOVERED
63d ago
2026-03-26
PUBLISHED
63d ago
2026-03-26
RELEVANCE
8/ 10
AUTHOR
sixteenpoundblanket