Agent memories are underrated

Kristoffer Berg — Fri, 11 Sep 2026 11:39:43 +0000

yes, if you are using claude it's most likely storing a bunch of memories. Very few knows this and even fewer knows what it is used for.

I've seen numerous power users on X and reddit struggle with these concepts. Mostly because they aren't motivated to sit down and read and adjust every single one.

Most people use claude.md/AGENT.md files and maybe agents and skill files. Since moving away from IDEs these files are more hidden than ever.

Lately I've been building all sorts of small apps that I can use to help me become more effective. I've built worktwin.io for running scheduled agents with access to more or less everything. It basically replaces Hermes, openclaw, grokbot and similar because I can create the agents in any ai tool.
For managing agents, skills and memories plus my knowledge graph I've been using heyrafter.xyz. The best advice I have gotten when building is to use it yourself. Up until now I have used both while I built it. But suddenly I had an epiphany

Memories can also be point in time data

Meaning I can have multiple session running and in one I might do a competitor analysis which I want to revisit. How do I do that? Create a task? No, because it's not something that I will do. Store it in a project repo? No, I want to access it everywhere maybe even in another AI tool.

Sessions are linear but we don't think linearly

So instead, I asked my ai agent to store it in Rafter. And from now on every session, every agent can pick it up and continue on this memory. Rafter essentially solves the linear problem.

So memories are underrated. They can be more actively used to share information between sessions and tools over time.

DEV Community: Kristoffer Berg

Agent memories are underrated