The system story — discipline, memory, hygiene, never-lost work, health, resilience

Keynote

wikiTaTa is a workspace where people and many AI coding sessions build software together. What makes it work is not one feature but discipline. Every session starts with the same rules, is held to them turn by turn, tracks what it spends, and writes what it learns to a card. The result is excellent memory management: both present memory (what a session holds right now) and recall (what any session can find later). Everything else follows from that.

Origins (March to May 2026)

• 16 March — the first cards, from day-to-day work on a pageant ballot app built with AI app builders. The pain was already clear: every new AI chat started blind and had to be re-told the project.
• 24 March — SGK ("SQL Git Keeper"): a web app with an AI wizard that connects to live databases, keeps secret keys out of code, logs every action and locks behind a PIN. Its value card says it plainly: it saves tokens by catching problems before expensive conversations.
• 25 March — the first behaviour rules: a session-start routine, directive rules, and working-style rules to use fewer tokens.
• 28 March — wikiTaTa is named, cloned from SGK.
• 4 April — the MCP server is specified: the database is memory, not chat history.
• 12 April — the multi-tenant user, entity and group model.
• 29 April — local models power hybrid search, and a second user connects.
• 30 April — the feature registry and "inventory first": ask what already exists before building.
• 29 May — the SvelteKit app starts, multi-tenant from the first commit.

The core ideas were all present in the first two weeks: save tokens, keep secrets out, start every session with rules, treat the database as the memory, wrap it in a desktop app. The months since have turned them into enforced, tested systems.

The parts of the story

1. Discipline — rules for how an agent behaves, set at session start and enforced every turn by hooks on every prompt and tool call.
2. Memory — every turn's tokens measured, a context-pressure warning, and the habit of writing anything worth keeping to a card.
3. Code hygiene — from one monolithic file to an optimised SvelteKit codebase, with duplicate implementations found and removed by gates.
4. Work is never lost — git, plus cards that always save, revisions and an audit trail.
5. System health — "am I running properly?" at every start, and a signed golden bundle that makes every seat the same.
6. Resilience — a primary server, a failover in another region and a third standby on another continent.
7. The desktop app — an Electron wrapper that can drive and guide the user.
8. The product surface — the canvas, everything-is-a-card, decisions answered inline, the Studio.

Significance: an AI session forgets everything when it ends. Discipline is what makes many forgetful sessions behave like one careful team.