This is a submission for the Hermes Agent Challenge: Build With Hermes Agent
What I Built
Every developer knows the feeling. You've ...
For further actions, you may consider blocking this person and/or reporting abuse
this is one of those problems that's 10% information and 90% accountability loop. most people in tutorial hell already know what to build - they just need the forcing function.
Exactly this. And I think that's why most learning tools fail they keep solving the 10% information problem with more information.
The Telegram nudge at 08:30 every morning isn't smart. It's not even that technical. But it's there, every day, with your name on it. That's the whole point.
The best mentor I ever had didn't teach me much I couldn't have Googled. He just kept asking "did you build it yet?" until I did.
yeah, the failure mode is never "not smart enough" — it is "smart but inconsistent." most habit tools personalize the content when they should be personalizing the pressure. your name at 08:30 > a personalized curriculum you check whenever motivation shows up.
"Personalising the pressure not the content" that's a better one-line description of Hermes Mentor than anything I wrote in the whole post. 😂
You've just described exactly why every personalised learning app with beautiful dashboards and AI-curated paths still loses to a friend who texts you "bro did you push today."
Might have to steal that line for the v2 landing page if that's okay with you.
haha steal it. the friend-who-texts case is the one worth solving for - social pressure doesn't compress into a feature, it's relational.
"Social pressure doesn't compress into a feature, it's relational" okay now you're just writing my thesis for me. 😂
And that's the honest limitation of Hermes Mentor right now. The 08:30 message shows up but it doesn't know you didn't sleep, had a bad day, or just got a new job. A real friend adjusts. The bot doesn't.
Maybe that's the actual v2 problem not smarter curriculum, not better gap detection, but making the pressure feel less like a cron job and more like someone who notices when you go quiet.
Hits the nail on the head! From scanning real repos to seamless Telegram integration, this is a true AI-native mentor. The local-first architecture with persistent local memory protects privacy while compounding development skills. A brilliant way to escape tutorial hell and revive stagnant code!
Thanks Andy!!
Great question! Entity disambiguation across sessions is actually where the trust scoring system shines. Instead of relying purely on LLM extraction (which can hallucinate entities), we layer it: (1) rule-based extraction for known patterns like names and project references, (2) jieba tokenizer for Chinese entity boundaries, then (3) a confidence filter that rejects entities with < 0.3 trust score. The unexpected win was the alias system — mapping "xulingfeng" ↔ "许凌峰" ↔ "许工" let us span English and Chinese mentions without duplicating entries. Curious — how does your approach handle multi-language entity references?
Glad if you post it in the same thread where we are discussing, nvm 😆
That layered extraction approach is really elegant rule-based first, then tokenizer, then confidence filter. The trust score threshold as a rejection layer is something I hadn't considered but makes total sense. LLM hallucinating entities into memory is exactly the kind of silent corruption that would be a nightmare to debug weeks later.
The alias system is genuinely impressive. Spanning English and Chinese mentions without duplicating entries solves a problem most Western-built tools don't even think about.
Honest answer on my side right now I don't handle multi-language entity references at all. The USER.md approach is flat markdown, so it's as smart as the LLM writing it. Works fine for English GitHub profiles but would fall apart fast with mixed-language repos or non-Latin usernames.
This is making me think the right v2 move is to not reinvent this and instead look at integrating something like what you've built as the memory layer rather than raw markdown files. A trust-scored entity store would make the audit results significantly more reliable over time.
Are you planning to open source MemBridge? Would genuinely love to dig into the implementation. 😄
Tutorial hell is real — and an agent that learns your progress instead of resetting every session is exactly the fix. Love the mentor angle.
Thanks Stephen!!
This looks super cool and useful! 👏🏻 One question, though: isn't the local run consuming too many resources? I remember running Gemma on Ollama got my laptop completely frozen at some point 🥶
Hey Klaudia! Your laptop won't turn into a space heater, I promise 😅
The LLM only fires once during the audit (~60-90 sec) then goes back to sleep. Daily nudges are just cached JSON , your CPU can relax.
Close Chrome's 47 tabs before running though. That's on you 😂
Ahahaha, you got me with this! 😂
Haha glad it landed! 😄
Let me know if you try it out 🚀
Love the Hermes Mentor concept — especially the "break the tutorial loop" angle. I've been running Hermes locally for a while and the execution boundary pattern you're using is solid.
One thing I'd be curious about: how are you handling the knowledge retention across sessions? Tutorial hell is partly a memory problem — you learn something, don't use it for 2 weeks, and it's gone. Does Mentor remember what you've covered?
Thanks! And yes memory across sessions was actually one of the first things I designed around, because you're spot on that forgetting is half the problem.
Every audit writes a
USER_<username>.mdinto~/.hermes/memory/which Hermes loads automatically at the start of every future session. It tracks gaps identified, current roadmap week, what's been closed, and audit history over time.So if you completed Week 1 and come back 2 weeks later, Hermes already knows it doesn't start from scratch.
The honest limitation right now: it tracks what projects you've pushed via GitHub re-audits, but it doesn't yet track conceptual retention (like "did you actually understand Jest or just copy-paste it"). That would need either a quiz layer or richer commit analysis. Definitely something I want to explore next.
Curious how are you handling long-term memory in your Hermes setup?
Great breakdown! The USER_.md approach is clever — we went a different route with MemBridge: SQLite with entity linking and trust scoring. The conceptual retention problem is real — curious if you have thought about integrating access frequency into the audit flow?
MemBridge sounds really interesting SQLite with trust scoring is a much more structured approach than flat markdown files. Would love to see how you handle entity disambiguation across sessions, especially when the same concept appears in different contexts.
On access frequency honestly hadn't thought about it that way but now I can't stop thinking about it. The idea of weighting audit signals by how recently and how often a language or pattern appears in commits is really compelling. A repo you touched once 2 years ago shouldn't carry the same weight as something you pushed to last week.
Could even go further commit density + file churn rate as a proxy for "am I actually using this or just copy-pasting." Richer than just language bytes.
Adding this to the v2 list for real. Thanks for the nudge this comment thread is turning into a roadmap 😄