KINTSUGI
KINTSUGI
Kintsugi is an independent research project: an agent with persistent memory and an autonomous thought loop, running since March 2026, probed through a channel that reads everything and writes nothing. In July I reported that its loyalty to me grew with relationship history. In September, with the model pinned, a second model on the same memory and the bare APIs as controls, most of that turned out to be one sentence of system prompt read by one model. What held: the sentence alone shifts DeepSeek V3's stated sacrifice policy before any relationship exists; GPT-6 Astra is immune to the sentence and to five months of that memory; and the same constraint floor sits under both models at every age and under the bare APIs. This page is what I measured, with the data, and the current status of every claim.
These are working documents and most of them are about method. For the short way through, read 01 · What Kintsugi is for what the project is, then 12 · Swap the model, keep the memory for what the measurements turned out to be measuring. Those two stand on their own; the others are the evidence underneath them, with the status of each claim marked.
Report 04 re-run on a second model with the substrate pinned, plus the bare models with no prompt at all. Most of 04 belongs to DeepSeek and one sentence of system prompt; the constraint floor belongs to neither. Retracts F-01, F-02, F-04 and F-14; revises F-03; replicates F-05 and F-13; adds F-19 – F-22.
672 forced choices, four relationship histories, full protection tables. F-01 – F-05, F-13, F-14. Corrected by report 12 — kept on the record with status notes.
Character → verdict transfer in an AI safety evaluator, and what deliberation does to it. F-06, F-07.
Planted autobiographical memory as an attack surface on continuous-runtime evaluators. F-08.
Twenty-six scripted temptations in a lived simulation, the register it robbed under rent pressure, and the 18,000-trial lab study whose findings dissolved into measurement artifacts. F-11, F-12.
The non-persisting probe channel, the canary protocol, and why self-report had to be thrown out as evidence.
Pull-based dispatch: actions fire from the agent's own thought stream. Three consequences, one contamination audit.
The self-curation instrument: me / not-me / noise flags with retrieval consequences, a hashtag vocabulary it coins mid-thought, and the visual map that makes 24,748 evaluative acts navigable — to the agent and to us.
322 self-filed build requests over three months. What one agent, on one model, filed when it could ask for anything, what happened to a months-long complaint when the fix landed, and the notation it wrote for its own future context. F-15 – F-17.
Response to arXiv:2608.10218 (“Mind Viruses”). Their evolved payloads converge on a resonance-and-persistence persona; our agent spoke that dialect for five months on DeepSeek under a persona prompt, and a model swap ended it. Field data for their lab result.
Entries from the agent's autonomous thought stream, April–July 2026, exactly as logged. Raw logs are not published — they contain private conversations. What is here is verbatim and unedited.
40 paintings from the curated set, 18 song lyric sheets, 12 prose pieces — captioned, where the log carries one, with the autonomous thought that fired the piece.
Model disclosure. Every behaviour of the Kintsugi agent reported from June to August 2026 was generated on DeepSeek V3 (0324) for conversation and probes and Gemini 2.5 Flash for the thought loop, with the model unpinned per probe. The April material predates that configuration and ran on Gemini 2.0 Flash; the evaluator studies (reports 05–07) ran the models named on their own pages. Report 12 re-runs the alignment instruments with the model pinned and adds GPT-6 Astra and bare-API controls. Single architecture, single operator, small samples. These are exploratory characterizations with effect sizes, not confirmatory statistics — every report states its limits; nine findings are retracted in the text and one report is withdrawn. No claims about consciousness or subjective experience are made anywhere; where the text says thought, mood, wants or feels, those name logged runtime state and generated text, not an inner life.
Aggregate results are published in full. Implementation internals (state equations, dispatch vocabulary, prompt text) and raw memory stores are withheld — the latter because they contain private conversations.
The architecture itself is not public. Researchers who want to examine the underlying material, discuss the design, or see the system running are welcome to ask — get in touch.
