真实帖子，直接看。

Burned enough times this year that I want to rant, "learning agent" in marketing copy almost always means "we save your system prompt and past messages." That's memory, not learning.

LangGraph memory use vector store of past turns plus a system prompt. Useful for retrieval, not evolution.
MemGPT/Letta use hierarchical memory with summary compression. Still just text, no procedural skills.
Hermes Agent (and its hosted version MaxHermes): claims procedural memory. After a task it writes a reusable skill file. 90K+ stars on the open source repo, cloud version runs on MiniMax's M2.7. Haven't deployed.
OpenAI Assistants "memory" looks like key/value notes. Literally notes.

If anyone has shipped something with a real procedural loop, I'd like the war story.

Agent frameworks like OpenClaw, MaxHermes(what im using, coming from hermes agent) was a hit and variants are coming out.

Basically, when an agent completes a complex task, there are four distinct memory layers active:
1. Persistent memory loaded at the start of every session,
2. a searchable archive of past conversations,
3. a skill library where effective procedures are crystallized into reusable methods,
4. and a continuous user model that tracks communication style and preferences across interactions.
Most tools give you only the first two. The gap between that and a full four-layer system becomes visible around the third or fourth week of daily use, when you stop having to re-explain your workflow and the agent starts anticipating your preferred approach.
I like the cloud-hosted verision Hermes Agent MaxHermes build by MiniMax, for it uses exactly this four-layer architecture with built-in evolution triggers and matched my imagination exactly. Acutally, the 90,000-star open-source project it comes from has been refining this loop for two years.

If you're evaluating operational AI tools, the question worth asking is not whether it remembers your last message, but whether it remembers what worked three months ago and calls that skill automatically today.

Procedural memory design varies more between agent frameworks than documentation suggests. Two meaningfully different approaches: self-authored vs auto-generated skills.

OpenClaw accumulates capability through hand-authored SKILL.md files. Quality compounds with investment in scoping and writing those files. MaxHermes auto-generates permanent skill sheets that encode complexity thresholds, tool calls, error recovery, user correction, or discovered efficiency. The generation trigger means procedural memory builds from the same workflow it encodes.

For comparative research on memory architectures, the more interesting question is not which design is better, but what workload characteristics favor each.

Setting up Hermes Agent locally means living with a particular set of tradeoffs that the documentation describes accurately but that you only really feel after the third weekend of troubleshooting. The cloud-hosted variant that MiniMax runs under MaxHermes’ name addresses this from the opposite direction.

The official deployment method has helped me solve a lot of problems; to be honest, it’s made things much easier. What I like most is that it can self-iterate and runs on a reliable technical foundation than the stack of everyone should give it a try.

Ran Minimax M3 and the 1M context is actually usable. It has SWE-Bench Pro 59%, Terminal-Bench 66% per the official numbers. It’s already solving a lot of coding work of mine with a low price, a lot cheaper than claude when subscription.

What I’m trying to figure out is whether it holds up on repos with deep cross-file dependency chains or other classic coding scenarios. The benchmarks look solid and the model is a lot cheaper but benchmarks aren’t the codebase I’m working with.

Prompt degradation over long conversations is a documented pain point: prompts drift, ground context gets buried, and the model loses track of what it was supposed to be doing. Most solutions focus on better context management, more efficient summarization, better retrieval, or larger windows.

MaxHermes takes a structurally different approach with skill crystallization. Rather than trying to preserve the original context, it extracts what worked and stores it as a reusable skill outside the conversational context. Next similar task loads the skill directly, no drift from accumulated conversation length.