In October I mostly wrote about systems you can run on your own infrastructure. That showed up in agent memory wired to local LLMs, the practical limits of letting a model make decisions, and self-hosted alternatives for S3, SEO platforms, and newsletter services.
The other thread was Go: framework benchmarks, keeping goroutines bounded, and what the 1.27 upgrade actually breaks. Read together, the month is a set of notes on keeping the stack local, bounded, and upgradeable.
1 AI agents, memory, and decision boundaries
The month’s AI work centred on how much context an agent carries and where the line is between a model recommending and a model deciding.
- Hindsight for Hermes Agent: Self-Hosted Memory with Local LLMs — agent memory in PostgreSQL, with local Qwen3 embeddings and reranking through Ollama or llama.cpp, no cloud inference.
- Tiered Context: A Knowledge Hierarchy for AI Agents — layering agent knowledge by injection trigger, from always-on context down to explicit reference, with profile boundaries that keep domains apart.
- Decision LLMs: When a Model Should Decide — and When It Shouldn’t — a rubric for the five roles a model plays in decisions, routing each to the smallest competent model, and keeping decisions verifiable.
- Trending Open Source AI Projects on GitHub, October 2026 — a snapshot of what’s trending this month, grouped by job: assistants, memory, RAG, local inference, builders, and gateways.
2 Local models and the hardware they run on
Two shorter pieces on the state of local AI — what is actually confirmed about the next Qwen generation, and what the GPU behind it costs.
- Qwen 4 Is Coming: What’s Confirmed and What It Means for Local AI — the confirmed details, the 27B benchmark ladder, and how the new sizes fit consumer GPUs.
- RTX 5090 Price in Australia: Listings and History — RTX 5090 and 5080 pricing tracked from June 2025 to late October, with shop averages kept separate from individual listings.
3 Go: frameworks, concurrency, and 1.27
Three Go pieces covering the choice you make once, the patterns you use daily, and the upgrade you do on a schedule.
- Go Web Frameworks in 2026: net/http, chi, Gin, Echo, Fiber Compared — a benchmarked comparison of the main HTTP stacks, covering routing, allocations, and the fasthttp tradeoff.
- Bounding Concurrency in Go: errgroup, Worker Pools, and Backpressure — errgroup limits, semaphore channels, worker pools, and the goroutine leak that survives
errgroup.Wait. - Upgrading to Go 1.27: What Breaks, What to Adopt First — the encoding/json v2 engine gotchas, the stdlib uuid migration, and the GODEBUG changes that fail builds.
4 Self-hosting the web and infrastructure stack
The largest section of the month: self-hosted alternatives for the pieces of a web stack you usually rent, from SEO down to the VMs underneath.
- Self-Hosted SEO Tools and Platforms: Open Source Guide — crawlers, rank trackers, and MCP tools, with a look at which features work without paid SEO APIs.
- WordPress SEO Plugins Compared: Yoast to Local AI — seven plugins compared, including how to keep inference on your own server.
- Best Newsletter Platforms in 2026: Pricing, Hosting and SEO — Buttondown, beehiiv, Kit, Ghost, Substack, and others, compared on pricing, public archives, and ownership.
- Self-Hosted S3 Alternatives in 2026 (RustFS, Garage, SeaweedFS, Ceph) — native object stores versus filesystem-backed S3 gateways.
- Clone a VM in Virt-Manager with virt-clone and virt-sysprep — the full pipeline for cloning KVM guests: virt-clone copies the disk, virt-sysprep resets identity.
If one of these articles is useful to someone building AI systems, backend systems, infrastructure, or technical knowledge workflows, please forward this email or share the link with them.
Thanks for reading, Rost