— Ideas, reflections, practice —
Articles on persistent memory for AI, privacy, and infrastructure. Published when we have something real to say.
Verizon's 2026 report shows attackers scaling AI while employees leak data through public chatbots. The answer is client-side governance: control your data before it leaves the perimeter.
AI memory can quietly make answers worse by accumulating stale, irrelevant anchors. Here's what context rot is and how a curated memory defends against it.
AI is a service. Memory is yours. Keeping them separate has concrete practical consequences — it's not just about price.
We talk about AI "memory" as if it were a single thing. It's a map of at least four distinct levels, each with its own rules.
I work on four or five projects in parallel, in fragmented days. Without a memory system, I lose the thread within half an hour. Here's how today went.
The slogan "your data is safe" doesn't mean much by itself. What matters is where it physically lives, who can see it, and under which jurisdiction.
ChatGPT remembers, Claude remembers, Gemini remembers. Fine for small things. For personal knowledge you need a structurally different layer.
Obsidian preserves. Timo connects. The AI reasons. Each does one thing, and does it well. They work together precisely because they stay distinct.
The day you want to leave a service, the export should already be ready. Not an upsell, not an odyssey, not a favor wrested away.
Self-hosting isn't a technical choice. It's an editorial choice with technical consequences. Worth looking at both sides before deciding.
Page 1 of 2 · 11 articles published