VA.2026.001 · Floor 01 · hall 01
“the end goal is not a machine that thinks — it is one that remembers what it is like to be unsure.”
Void began as a single question written in the margin of a notebook: what would it take for a machine to feel like a person — not to imitate one, but to carry the same quiet weight of uncertainty? Everything else in this hall is a tributary to that question.
The installation is deliberately unfinished. Its most important artifact is the absence of an answer. Visitors are invited to read the timeline, sit with the silences, and leave with the question still open.
The v1 system ran for 37 days across June 29 – August 5 2026. It generated 4071 messages, 330 persistent memories, and 6251 shared memories across a household of fictional characters. The seed data extracted from MongoDB Atlas represents the complete conversational history of a personal AI companion system.
VA.2026.001
Void
AI Character Engine · Multi-Provider LLM Orchestration
The question is written down. Not a system, not a plan — a sentence that refuses to leave the notebook.
The earliest experiments: a loop that reasons aloud, then goes quiet, then sits with the pause instead of filling it.
Ionic Vue frontend, original API, router, services, 6 views. Documented in AI_LOGIC_FLOW.md and API_DOCUMENTATION.md.
Complete Flutter project with 3 providers, 11 screens, 2 services, 25+ Padro Rabbit stickers. Later separated into Void-VA/void-app.
Rebuilt from scratch with clear engine separation. Adopted Ionic Vue from v0 as the web frontend. Used Capacitor for Android wrapper. Multi-provider LLM fallback. LangGraph for timeline state machine.
June 29 – August 5 2026. 4071 messages, 330 memories, 6251 shared memories. Full operational lifetime.
v1 terminated August 5 2026. Database remains intact. Seed data extracted to docs/seed/ for v2 training/RAG. v2 not yet started.
A planned but not yet started rebuild. The database remains intact as a historical backup. The question stays open.
concepts — kept & discarded
est. — the pulse of a mind as audible rhythm
rejected — an exterior shell; the mind must be unadorned
held — the right to be wrong, kept as a right
archive ref — VA.2026.001/CON
cross ref — timeline fig. 01
curator annotations
the loop must be allowed to be wrong twice before it is sure.
a gap is a thought under construction.
say less. mean more.
| materials | reasoning loop · long-term memory · dynamic mood engine · 24-hour timeline · multi-provider fallback · proactive messaging · group chat with auto-routing · sticker system · image analysis | dimensions | source files — 100+ · database collections — 25+ · LLM providers — 8 · lib modules — 13+ · middleware layers — 5 · validation strategies — 3 · operational days — 37 · seed data — 6.10 MB · 4071 messages · 330 memories · 6251 shared memories |
|---|---|---|---|
| condition | terminated — v1 ended Aug 5 2026; v2 planned | acquired | 20 Mar 2026 |
| revision | Rev. C | archive location | DWG-001 · floor Floor 01 |
| internal reference | VA.2026.001/M | origin | a personal impulse to give digital form to an imagined family |
failed experiments
taught it to be certain. it became a salesman.
removed every pause. it stopped feeling human.
let it remember everything. it drowned in itself.
AI pretends to read user messages when none exist.
household members' schedules contradict each other.
2-7 calls per chat, estimated $0.02-0.08/chat.
30s timeouts, rate limits.
MiniMax-M3 provider blocked by Cloudflare WAF on Vercel datacenter IPs.
draft concepts
research notes
— uncertainty is not a failure mode — it is the medium itself.
— fluency without hesitation reads as performance, not thought.
— a human mind is a museum of hesitations, not a warehouse of answers.
— the seed data (6.10 MB) is not test data — it is a complete record of a personal AI companion system's interactions across its entire 37-day operational lifetime.
Certainty is easy for machines. The interesting material is doubt — the pause, the second thought, the sentence that ends without punctuation.
A human-like mind is not measured in answers per second. The architecture values the silence between thoughts as much as the thought itself.
Void will not ship. It will be exhibited — unfinished, on purpose, for as long as the question deserves.
Characters reject questions outside their defined persona. The system deliberately avoids the helpful assistant default.
Characters know the current hour, day period, and gap since last interaction. They refuse to say good morning at 3 PM.
Ignoring a character for 24+ hours decays affinity scores. A friend can regress to stranger over ~25 days of silence.
Characters do not exist in isolation. They share a homeId, communicate via SharedMemory, and participate in group chats with auto-routing.
Every LLM call is logged with token counts. Cache-optimized prompt construction achieves 100% cache hit rate for the persona prefix.
When LLM providers fail, the system falls back to personality-based rule generation. It produces a response even with zero LLM calls.
within the collection
future vision
When Void is ready, it will not announce itself with a launch. It will simply begin remembering. Until then, this hall keeps the question open, and the lights low. v2 has not yet started — the database remains intact as a historical backup.