A small number of essays we actually stand behind: how to think about LLMs, why memory needs an architecture, and why governance is the layer everyone skipped.
Every coding agent starts every session from zero. Guilds gives each of your projects a named agent with memory of its own, and the difference is the first thing it says.
Read Full ArticleSetanta — guild memory loaded: 3 open tasks, last worked yesterday
I'm Setanta, this project's agent, carrying the memory of every session we've had here.
Where we left off:
· Programme visibility shipped and verified on prod
Field notes from building the platform, not a content calendar
Deleting a memory is easy. Erasing a person from a system that has embedded, summarised, cited and logged them is not. What our purge destroys, what it keeps, what it scrubs from the logs, and the one thing it says it cannot touch.
An AI answer is a new document made from old ones, and it needs a wall of its own. Ours inherited the union of its sources' owners, the wrong set. What the rule is now, in three parts, and why the last part is the asker alone.
Before law, an authority expert over 154,000 passages of the Church's documents: graded, superseded rather than faded, with a citator. Law's exact reasoning shape, none of its licensing walls, and where the engine under Irenaeus was proved.
We test our authority-expert by trying to break it: 21 adversarial probes, a documented criterion for each, three runs through the public edge. The harness, its history including the bad days, and today's number.
Whose version wins, who may see it, does it still stand: the hard problems of AI memory were solved centuries ago by courts, libraries, and one small town in Donegal. On building memory by translating institutions.
The same Claude Code sessions run twice on our own production repo, once with guild memory and once without. Real transcripts, honest ties, and the three kinds of knowledge that never reach a repository.
Out of the box versus out of the box: 62.4% to 48.8% under one strict judge, with the full method, the fairness ledger, and the null results published alongside the wins.
The essay that started Yohanun: why every conversation with AI feels like starting over, and how a memory runtime solves the problem that bigger models never will.
The biggest architectural mistake teams make with AI isn't technical; it's conceptual. Why treating LLMs as specialized compute, not complete systems, changes everything you build.
Where traditional application patterns work well for AI, where they break down, and why intelligent systems need an infrastructure layer MVC never anticipated.
These essays are opinions held with reasons. If you're building in this space and see it differently, or the same, we'd genuinely like to hear it.
Email Us →