History Is Not Memory: Stake, continual learning, and persistent AI agents

The preprint of History Is Not Memory: Stake, continual learning, and persistent AI agents, with the verifiable records that stand behind every number in it. Persistent AI agents keep memories, update their weights and carry self-descriptions from one session to the next, which makes them look historical. The paper asks what that history is made of, and measures it. Replace an agent's accumulated state with an equivalent copy and record the residue, defined against the noise of an independent replay: a short proposition shows that an exact copy of a digital state leaves zero residue whenever two assumptions hold — the state is completely enumerated, and the computation is the same wherever the copy runs. The measurement then runs, under a preregistration and two independent audits, on a language agent learning continually over 42 sessions with five simulated users. Its past sits in two carriers, memory and weights, and it is path dependent: after a preference reversal the old preference resists the new one. Halting the agent, resuming it in a new process or copying it left no residue, down to the last stored bit. Copying its memory alone lost most of what it knew. Its account of itself followed the memory file: with only its weights it acted on preferences it could not report, and with another user's memory it described that user's preferences as its own. Two registered expectations failed and are reported as registered. For agents built this way, history is a record. Stake — the mode of persistence in which a system's own activity produces its continued existence — cannot be read off a record; the pause test settles it, and here it finds nothing at stake. The paper relates this to memory-based and psychological-continuity theories of personal identity and states five predictions, each with the outcome that would refute it. Files Yuan_2026_History_Is_Not_Memory.pdf — the preprint (15 pp.). HINM_v1.0_records-bundle.zip — the records: the run tree's checkpoints, probes, manifests and logs (1,170 files), the aggregate summaries and the numbers manifest, the two independent audits with their provenance and transcripts, the preregistration and its amendment register, and the tracer that walks each number back to a signed line in the preregistration. Start here python tools/trace.py --file results/main/aggregate/paper_numbers.json --all Run on the complete tree (weights present) it reports: 61 key(s): 61 intact to a signature, 0 intact but unsigned, 0 broken OK: every chain reaches a signed registration. That output is included unedited as TRACE_GATE_COMPLETE_TREE.txt. On this record the same command stops at level 5, because the 113 LoRA weight files (*.safetensors) are not archived — deliberately, and stated in the paper. Their SHA-256 are recorded in audit-1/preaudit_runs_sha256.txt, so a copy of the weights obtained elsewhere can be checked against this bundle. Every other level (L1-L4, L6-L8) verifies against the files here, and TRACE_GATE_ON_THIS_BUNDLE.txt shows exactly that. The experiment sources at the commits that produced and analysed the runs (src/**, tools/**, config/**) are included under source/, exported with git archive and byte-identical to those commits, so SOURCE_DIGEST can be recomputed and checked against the value pinned in the preregistration; SANITIZATION.md, section "The source snapshots are not sanitized", records this. EVIDENCE_INDEX.md records what is and is not here. MANIFEST.json lists every file with its size and SHA-256. SANITIZATION.md lists every file whose text differs from the private original, with the hash before and after: machine paths and the author's given name were replaced in 13 files; no hash value, score or log line was changed.