These are my "LLM Wiki" notes, I've been using to build a central resource of "how to build an AI First Startup"
If you've got any other tips *I'd love to hear them* - thank you!
Over the last few days I've added peer-to-peer communication, using a simple message passing and status checks... no need for complicated hooks etc.
One repo agent found a bug and told the curator repo agent what it was, and it went ahead and fixed it
Over the last few months I've been quietly adding more features to elfmem my "evolving" agent memory library ( https://github.com/emson/elfmem ). I'm now able to use it in Claude Code et al. to curate my repos, and even write indie articles based off my obsidian vault. Anyway this article explains how. I'd love a comment, or even a github star! Thanks Ben
This has been a fascinating project. I've had the chance to explore some really original ideas. Things like: blocks that can be calibrated, mind blocks that can be used to simulate other agent behaviors, graphs, dreaming, frames that return a SELF or ATTENTION, etc. group of memory blocks, and peer to peer message passing.
All this has led to a really powerful agent framework you can use today.
I'd love a comment... thanks! Ben
Agents need memory that evolves. Memory blocks should degrade or get promoted over time, this opens up agents as repo curators... magic then happens if they can then talk to each other.
This is interesting as there is definitely a middle ground for agent memory. On the openclaw side you have a single MEMORY.md file on the other you have RAG and GraphRAG.
I wonder if Agent memory should be more nuanced? When an agent learns something how should it promote or degrade these memory blocks - you don’t want a trading agent memorising a bad trading pattern, for example.
Also the agent might want to recall semantically similar memories, but it might also want to retrieve block relationships or groups of blocks for different purposes.
We’ve been exploring all these concepts with “elfmem” (sELF improving MEMory): https://github.com/emson/elfmem
Would love your feedback!
Thanks so much for the comment! Yes, I initially had LiteLLM gateway and backed it out after the security issues.
Memory is becoming a crucial part of agents and there isn't a one size fits all solution - unfortunately. I found I wanted to replicate things like SOUL.md but in a way that could "evolve", so this project has "frames" which is just a collection of tagged / filtered memory blocks which can be used to identify specific concepts.
It also allows the agent to use "outcomes" to calibrate it's own memory blocks, which is really powerful.
Anyway I think "cognitive loops" are going to become more relevant, and will be an art. Thanks again... let me know how you get on!!
Good question. I've had some success with Qwen2.5-Coder 14B, I did use the quantised version: huggingface.co/Qwen/Qwen2.5-Coder-14B-Instruct-GGUF:latest
It worked well on my MacBook Pro M1 32Gb. It does get a bit hot on a laptop though.
This is really useful I often find I want to spin through different designs and I was wanting something like this for my own projects.
Have you thought about being able to go back to previous designs?
Would also like to Shadcn components in the mockup too.
Really cool. Thanks!
reply