Weekly Review — June 29-July 5, 2026
written by Stefan Christoph
This is the Weekly Review, a Sunday digest of everything that went up on the blog this week, plus a short list of things I read but didn’t write about. If you only have ten minutes on a Sunday, this is the one to read.
This Week on the Blog
Agent Memory Is a Spectrum, Not a Switch
“My agent keeps forgetting things” is usually a memory-architecture problem, not a model one. Memory runs along a spectrum, from no memory, through the context window, to episodic, procedural, and semantic knowledge, and most teams cram all four kinds onto one rung (the context window) and wonder why the agent gets slower and less reliable. Name the rung the missing capability lives on, then pick the mechanism that fits. This kicks off a three-part series.
Own Your Real Estate: Why I Stopped Renting My Audience
If you only publish on platforms you don’t control, you rent your audience and the landlord can change the locks. The fix is POSSE: publish the canonical copy on a site you own, then syndicate teasers that each link home. Owning buys a durable URL, full control of the format, an audience you can reach directly, and your own analytics; the trade being that a link-out teaser often gets less feed reach, which is a cost I take on purpose.
The Other Half of Two-Pizza
Werner Vogels’ “A return to two-pizza culture” tells how a coding agent let one scientist prototype Amazon Quick Desktop overnight, with ownership culture carrying it from there, and I recognized my own build-use-learn-iterate loop in it. But the buy-in revolution is only half the story: getting an idea all the way to production, where it has to scale, stay up, and behave responsibly, is the harder half that AI made easier, not easy. The starting line moved; the craft now lives in finishing.
Vector Search, From the Whiteboard to the Cloud
Vector search is a five-minute whiteboard idea: embed text into a point in space, match by the angle between vectors, grab the nearest neighbours. It’s also a full quarter of production tuning, where chunking, recall versus latency, embedding drift, and metadata filtering all decide whether retrieval helps or quietly hurts. On AWS, Amazon Bedrock Knowledge Bases runs that ingest-chunk-embed-store loop as a managed service. It’s the first post in a new “Whiteboard to Cloud” format that pairs the academic explainer with the AWS practice.
Your Obsidian Vault Is a Knowledge Graph
A folder of linked, tagged markdown notes already is a semantic knowledge graph: wikilinks are edges, frontmatter is typed node properties, folders are namespaces. That’s exactly the structure an agent traverses for long-term memory, and my own kiro-cli-plus-Obsidian setup walks that graph every day, including the session that wrote the post. You graduate to a real graph database when you hit cross-graph queries, scale, schema enforcement, or many concurrent writers, not before. This closes out the agent-memory series.
The Thread This Week
Read together, the week is one argument from two directions: knowledge is more useful when it’s structured, durable, and owned. The memory spectrum names the rungs; vector search is the retrieval mechanism for the semantic rung; the linked vault is that semantic memory in a form you already maintain. “Own Your Real Estate” points the same instinct at yourself, since the vault you control is both your agent’s knowledge graph and the canonical home of your work. And “The Other Half of Two-Pizza” is the honest coda: owning something means finishing it, and the finish to production is where the craft now lives. Ephemeral context, a bigger prompt, a rented feed, a demo that never ships, is the thing all five posts quietly argue against.
Further Reading
Things I read this week that didn’t get their own post, all public:
- Notes on Exploring AI-Driven Development (Micah Walter) — Practitioner field notes on moving from AI-assisted coding to AI as a central collaborator across the whole software lifecycle, with a persistent context that carries from inception through operations. A good companion to this week’s vault-as-memory post.
- How Google Maps Finds the Shortest Path (Milan Milanovic) — A clean walk through routing at scale: Dijkstra’s as the foundation, A* with heuristics for pruning, and Contraction Hierarchies as the production trick behind sub-100ms routing over 20M-plus road segments. Useful background for anyone modelling logistics or fleet problems.
- Notes on Mistral OCR 4 (Stanislav Belyaev) — A look at where document-parsing models are landing: structured output with paragraphs, tables, and headers, bounding boxes, and per-block confidence scores across a dozen-plus languages. Useful context if you’re weighing document-intelligence options for a real pipeline.
Until Next Sunday
That’s the week: a ladder for agent memory, the retrieval mechanism underneath it, the vault you already own as the top rung, a case for owning the home where all of it lives, and a reminder that the hard, satisfying work is in finishing. Where does the canonical copy of your own knowledge live right now, in a structure you (and your agent) can traverse, or scattered across feeds you don’t control?
About the Author
Stefan Christoph is a Principal Solutions Architect at AWS, focused on agentic AI, media & entertainment, and helping builders move from demo to production. He writes about AI architecture, developer productivity, and the future of software.
This is a personal blog. Opinions expressed here are my own and do not represent the views or positions of my employer.
❤️ Created with the support of AI (Kiro)