The Agent Is Its Record: exact copies, no hidden carrier, and what writing to memory does to a continually learning language agent
Year: 2026
Type: preprint
Abstract: A language agent that keeps a memory and retrains its own weights looks as if it carries a history. I test, under signed preregistrations, the claim that this history is nothing over and above a finite record: a memory window, a weight adapter, a training buffer and a few counters. If the claim holds, two things follow. Copying the record copies the agent exactly, and writing into the record changes the agent directly. Across seven preregistered experiments on Qwen2.5 agents (1.5B and 0.5B parameters) that learn over 42 sessions with five simulated users, all 144 comparisons between a restored or independently replayed agent and the agent it copied were bit-identical. Twelve perturbations that the record does not contain, among them a different hash seed, a killed and restarted process, a machine reboot, a moved system clock, a RAM-backed disk and heavy CPU contention, changed no stored bit and no probe margin, while three positive controls showed that the instrument responds to a change. A preference that had left the memory window and never reached the weights fell to the level of a fresh agent. On a different CPU and operating system the agent kept every one of its decisions, but not its bits: margins moved by up to 0.0002 and the adapter diverged after retraining. I then used exact copies to run clone-and-edit experiments. Placing a short user summary before the memory window weakened the agent's hold on its recent preferences in all five users (by 0.79 log-odds), and a more specific summary weakened it more. Writing one contradicting entry per preference into memory flipped about a quarter of the agent's choices; preferences already consolidated into the weights moved less than those held only in memory, in all five users. Two registered predictions could not be decided and are reported as such. For agents built this way, the past is a record, nothing in them is at stake in the sense of the pause test, and memory features and memory attacks are both edits to that record. Files: the paper (15 pp.) and the reproduction package, containing per-item records of all modules, signed preregistrations, revisions, rulings and deviation registers, and the source code of every run. "python reproduce/recompute_paper_numbers.py" recomputes every number in the paper from the records (Python standard library only, about 2 seconds). This is the third paper in a programme: The Anchor Paradox (doi:10.5281/zenodo.23029149) proposed the theory; History Is Not Memory (doi:10.5281/zenodo.23170703) ran the first worked example.
Author: Yuan, Simin
Language: English
Subfield: History and Philosophy of Science
Field: Arts and Humanities
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7223246107