VA.2026.015 · Floor 01 · hall 15
“the engines were never the exhibit; the moment they agree with each other is.”
VA-Prototype is not a new engine — it is the room where the engines meet. Identity, memory, chat, graph, maps, and the day planner are mirrored in as local packages and driven by one process. The point of the exhibit is not any single mechanism but the seams between them: where recall meets voice, where a plan meets a tick, where one persona's day collides with another's.
It is deliberately a single runtime rather than a published platform. Packages are copied in and synced from their source repositories, so the prototype can move faster than releases and reveal interface lies early. The cost is honest: copies drift, and the drift shows up as build errors rather than as a clean integration story.
Time is the load-bearing feature here. Conversation carries the gap since the last exchange, plans are written once at rollover and only looked up afterwards, and a cron advances the simulated day in five-minute steps. Most of the late revisions were not new capabilities but corrections of when the machine believed it was — and the difference between what is happening now and what merely happened before.
VA.2026.015
VA-Prototype
Persona Runtime · Integration Prototype · Cron-Driven Ticks · Elysia + Bun
Elysia + Bun skeleton, module registry, and the engine packages mirrored into a single process. Nothing new invented — only placed next to each other.
The day planner is wired into the tick pipeline; chat begins to follow the day's plan and the persona's profile instead of improvising every hour.
Persistence moves to a SQL-over-HTTP gateway, with writes batched into one request. The prototype gains a second driver so it can fail over between transports.
Session tails are read back so a conversation can resume across gaps; recall gains a recency fallback; plan refresh is collapsed to one request per household.
Now and then are separated in both chat and memory; a single temporal context replaces token-level gates; tick advance becomes several times faster through batching and fewer model calls.
concepts — kept & discarded
kept — one process, all engines; integration before publication
rejected — releasing every package separately to prove maturity
held — a mobile surface, drawn but not built
archive ref — VA.2026.015/CON
cross ref — timeline fig. 01
curator annotations
a system is legible at its seams, not at its features.
copying a package is a loan against tomorrow's build.
the machine may forget; it may not pretend the past is the present.
| materials | multi-persona registry · persona chat runtime · persistent memory capture · identity resolution · personality-based day plans · tick advance cron · temporal context · SQL-over-HTTP driver · seeded personas · integration notes | dimensions | engines wired — 6+ · commits — 82 · crons — 2 · database drivers — http · postgres · integration days — 16 · version — 0.1.0 |
|---|---|---|---|
| condition | prototype · v0.1.0 · 82 commits · six engines mirrored in · two crons live | acquired | Oct 2026 |
| revision | Rev. A | archive location | DWG-015 · floor Floor 01 |
| internal reference | VA.2026.015/M | origin | assembled from the parts of a discontinued companion, Oct 2026 |
failed experiments
copies of the engine packages fell out of step with their sources; builds broke where behaviour had quietly changed.
the cron spent its budget on round trips; batched writes and fewer model calls brought it down sharply.
a reply could be captured under an identity that already existed, leaving a duplicate with no speaker. Assigned a new sequence and logged instead of dropped.
draft concepts
research notes
— integration is the cheapest test of an interface — the disagreements appear in hours, not sprints.
— mirroring beats publishing while the contract is still moving; the drift is the feedback.
— a prototype should be embarrassing at the joints and presentable at the edges.
A library that has never been consumed by another library is a hypothesis. The prototype consumes all of them at once so the contradictions surface early — and the fixes land where they belong, in the engine or at the seam.
Simulation is only believable when every participant agrees on what time it is. One runtime, one advance, one shared notion of now — no engine keeps a private calendar.
Every tick and every reply has a budget. Plans are written once per day, writes are batched, and refresh is bounded — not because the work is unimportant, but because an expensive simulation never runs long enough to become interesting.
within the collection
future vision
The prototype stays unbeautiful on purpose. A mobile surface sits in the notes, the mirror sync waits for the contracts to settle, and the next revision will be judged by one question only: do the engines still agree when no one is watching them?