Weekly Roundup #29 — RAG, Agent Memory, and Homelab Wins

This week was all about testing assumptions. I spent most of July 7-12 running experiments on my own stack — RAG retrieval pipelines, homelab logging, and agent memory architecture. The kind of work that doesn’t look flashy in a demo but changes how your system behaves under load.

I published 4 posts this week across AI engineering and infrastructure. The numbers are modest — this is a builder’s blog, not a media site — but the engagement signals are telling me something useful.

The Numbers

4 posts published. Topics spanned RAG retrieval, agent memory, homelab optimization, and a weekly roundup. The category mix was heavy on practical AI engineering — no fluff pieces, no hot takes on the latest model release. Just things I actually built and broke this week.

Highlight: Stop Building Agent Memory

The post that surprised me most was “Stop Building Agent Memory — Your Agent Doesn’t Need It.” It pulled 3 views with a 22-second average time on page and a 0.3 bounce rate. That’s the best engagement I’ve seen all week. People who clicked through actually read it.

The argument is simple: most agent memory implementations are premature optimization. If your agent runs in a stateless loop with context windows that fit the task, you don’t need a vector store or a memory module. You need better prompt design and a cleaner tool interface. I’ve been guilty of over-engineering this myself, and the post came from a real refactor where ripping out the memory layer made the agent faster and more reliable.

The other strong performer was “I Tested 5 RAG Strategies. Only 2 Worked.” with 4 views. That one resonated because it’s honest — most RAG tutorials show you the happy path, not the three failed approaches you try before finding something that works. The two that worked were hybrid search with reciprocal rank fusion and a carefully tuned chunking strategy based on semantic boundaries rather than fixed token counts.

What’s Next

Next week I’m diving deeper into the RAG pipeline that actually worked — full architecture, code, and the benchmarks that convinced me. I’m also working on a comparison of local LLM serving options for homelab setups, since the logging post got enough traction to warrant a follow-up on the infrastructure side. And I’ll keep the agent memory debate going with a practical guide on when you actually do need it (and how to build it right when you do).

Thanks for reading. Build something real this week.

Related reading: my hands-on testing of 5 RAG retrieval strategies and why most agents don’t need persistent memory.