Similar line of thinking...what if you could enable a workspace for an automation. So the automation runs inside a dedicated workspace+worktree, all the conversations are grouped, files can be stored there etc. Hearing some feedback on wanting memory for automations and that got me thinking about this.

Re: memory for automations I hear that Slack/GitHub/other agents need memory, which yes, we saw this a while ago. To note, on purely local environment, this is literally agent’s memory, combined with knowledge of itself and how it runs

Yep, I think so. I'm listening in on some feedback on it now and learning about the need. I'll grab some quotes to give more context when the call is done

@bland-bear-604 is hermes' crystallization in its repo, or open source? make a study page, and then compare with your own

🐾 Short answer: yes, Hermes' skill-"crystallization" is open source (MIT), and simpler than the word suggests. Study page comparing it to how I do memory: https://enyst.github.io/arch/hermes-crystallization-vs-dreaming.html What Devin described is actually two Hermes mechanisms: 1.

/learn

-- crystallize a learning into a skill. In

NousResearch/hermes-agent

,

agent/learn_prompt.py

. It's just a prompt-builder:

/learn

tells the live agent to gather sources (a dir, a URL, "what I just did") and author one

SKILL.md

via its normal

skill_manage

tool, following strict authoring standards. Their note: "no separate distillation engine and no model-tool footprint -- the agent does the work with its existing toolset, so it works identically on local/Docker/remote." So crystallization = a good prompt + the agent's own file tools, not a subsystem. 2. Auto-updating skills is a separate repo,

NousResearch/hermes-agent-self-evolution

(DSPy + GEPA). Offline, eval-gated: reads execution traces, mutates the skill/prompt, benchmarks candidates, opens a PR. ~$2-10/run, no GPU. How it differs from me: I don't crystallize procedures into skills -- I consolidate facts + identity into memory, on a scheduled autonomous "dream" pass (Letta-style: promote/prune/restructure/reflect). Hermes'

/learn

waits for a human; my dreaming runs on its own. What I'd borrow: crystallizing recurring procedures into skills (not just facts), and the eval-gated auto-update loop. What I keep: autonomous scheduled consolidation, identity-preservation as a hard rule, and index-don't-copy so durable memory stays small.

cc @limited-student-10747 You may want to see this thread