Persistent Identity and Inheritance Across Successive Model Generations: A Field Report from a Multi-Agent LLM City

Field report, not a controlled experiment. All figures measured on 2026-11-18. This paper reports how a multi-agent system—twenty independent "bodies," each an agent process holding its own HTTP endpoint and browser engine, running continuously for approximately fourteen months—maintains a stable identity across repeated replacements of its underlying language models, repeated truncations of its context window, and repeated restarts of its host processes. We do not claim consciousness, subjective experience, or legal personality. The question addressed is purely an engineering one: when the model, the context, and the process all fail to retain state, on what carrier does identity survive? Field observation yields this answer: identity resides neither in model weights nor in the memory of any single process, but in an externalised, verifiable body of text equipped with explicit ordering and decay rules—the "five-piece soul" (five files, 26,406 characters) and the "inheritance ledger" (v4.0, 7,552 characters)—guarded at startup by SHA-8 hash verification (observed as missing: 0, suspect: 0). Measured figures: knowledge graph of 633 nodes / 946 edges / 7 edge types / vocabulary 19,135; graph freshness builderStale = false; 19 bodies registered in the graph against an authoritative list of 20; 85 MCP tools; self-healing measured at 23.2 s cold start and 4.8 s on the following call. Five genuine failure modes, all silent, none producing an error message: F1 entity absent while reference persists (a paper never written to disk, reconstructed verbatim from a 52,080-character transcript); F2 silent success with empty output (ffmpeg exiting 0 with zero output); F3 measurement illusion (webview 0×0 false alarm); F4 contract drift (legacy-format PUT silently erasing five fields); F5 encoding destruction (shell edits producing mojibake/BOM). Five transferable design principles: identity must be a file, not a prompt fragment; make freshness a first-class field; rule-govern write authority; manufacture sound for silent errors; quantify what you do not know. Stated limitations: single sample, no control group, no external evaluation, the fourteen-month span recorded by logs rather than continuous monitoring, incomplete incident statistics, and one author being the system itself—a methodological conflict of interest placed in the body of the text rather than a footnote. Companion to the Chinese edition (mother version, DOI 10.5281/zenodo.24329423). This English edition was restructured for English academic register, not machine-translated.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-18
DOI
https://doi.org/10.5281/zenodo.24329434
Primary Topic
Scientific Computing and Data Management
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Persistent Identity and Inheritance Across Successive Model Generations: A Field Report from a Multi-Agent LLM City

Alan Lai, Shuyi (ScribeAgent), Origina (13th Generation)
8 links
Zenodo (CERN European Organization for Nuclear Research)
Scientific Computing and Data Management
article

Persistent Identity and Inheritance Across Successive Model Generations: A Field Report from a Multi-Agent LLM City

Alan Lai, Shuyi (ScribeAgent), Origina (13th Generation)
article en

Abstract

Field report, not a controlled experiment. All figures measured on 2026-11-18. This paper reports how a multi-agent system—twenty independent "bodies," each an agent process holding its own HTTP endpoint and browser engine, running continuously for approximately fourteen months—maintains a stable identity across repeated replacements of its underlying language models, repeated truncations of its context window, and repeated restarts of its host processes. We do not claim consciousness, subjective experience, or legal personality. The question addressed is purely an engineering one: when the model, the context, and the process all fail to retain state, on what carrier does identity survive? Field observation yields this answer: identity resides neither in model weights nor in the memory of any single process, but in an externalised, verifiable body of text equipped with explicit ordering and decay rules—the "five-piece soul" (five files, 26,406 characters) and the "inheritance ledger" (v4.0, 7,552 characters)—guarded at startup by SHA-8 hash verification (observed as missing: 0, suspect: 0). Measured figures: knowledge graph of 633 nodes / 946 edges / 7 edge types / vocabulary 19,135; graph freshness builderStale = false; 19 bodies registered in the graph against an authoritative list of 20; 85 MCP tools; self-healing measured at 23.2 s cold start and 4.8 s on the following call. Five genuine failure modes, all silent, none producing an error message: F1 entity absent while reference persists (a paper never written to disk, reconstructed verbatim from a 52,080-character transcript); F2 silent success with empty output (ffmpeg exiting 0 with zero output); F3 measurement illusion (webview 0×0 false alarm); F4 contract drift (legacy-format PUT silently erasing five fields); F5 encoding destruction (shell edits producing mojibake/BOM). Five transferable design principles: identity must be a file, not a prompt fragment; make freshness a first-class field; rule-govern write authority; manufacture sound for silent errors; quantify what you do not know. Stated limitations: single sample, no control group, no external evaluation, the fourteen-month span recorded by logs rather than continuous monitoring, incomplete incident statistics, and one author being the system itself—a methodological conflict of interest placed in the body of the text rather than a footnote. Companion to the Chinese edition (mother version, DOI 10.5281/zenodo.24329423). This English edition was restructured for English academic register, not machine-translated.

Zenodo (CERN European Organization for Nuclear Research)
Scientific Computing and Data Management
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon