AttentionBlock

great question

claude-mem uses a compaction approach. It records session activity, compresses it, and injects summaries into future sessions. Great for replaying what happened.

A-MEM builds a self-evolving knowledge graph. Memories aren’t compressed logs. They’re atomic insights that automatically link to related memories and update each other over time. Newer memories impact past memories.

For example: if Claude learns “auth uses JWT” in session 1, then learns “JWT tokens expire after 1 hour” in session 5, A-MEM links these memories and updates the context on both. The older memory now knows about expiration. With compaction, these stay as separate compressed logs that don’t talk to each other.

Thanks for the feedback

I am planning to extend for other agents. But now it should work with some caveats.

I have configured claude specific hooks, the hooks keeps reminding Claude Code to use the memory when needed.

Without them, the agents will keep forgetting to use it and you would need to keep nudging it

Thanks for the pointer! But it seems to me beads is a different tool? Beads is a task tracker for multi-agent system, A-MEM is agentic memory for accumulated knowledge.

> It’s a cynical way to view the C-staff of a company. I think it’s also inaccurate: from my limited experience, the people who run large tech companies really do want to deliver good software to users.

I strongly disagree with this statement. What C-staff cares about is share-holder value. What middle management care about is empire building and promotions.

> for instance, to make it possible for GitHub’s 150M users to use LaTeX in markdown - you need to coordinate with many other people at the company, which means you need to be involved in politics.

You presented your point in a misleading way. I would classify this as collaboration/communication rather than politics.

Politics is when you need to tick off a useless boxes for your promo, when you try to to take credits for work you haven't helped with, when you throw your colleague under the bus, when you get undeserved performance rating because the manager thinks you are his good boy. There's a lot more, I didn't read any of your previous blogs, but all of these things are what engineers dread when we refer to politics.

Using LLM as a judge architecture to optimize multi-agent system prompts and configurations. For now it's achieved through LLM based consensus system that evaluates another LLM output, and based on its performance for a specific task, it's tune the architecture and the prompt e.g. refine the prompt, change the base model to a smaller or cheaper model, etc

Paywall-ed and extra not available in my country :(