A Model-Independent Memory Substrate: Preserving an Agent’s Account of a User Across LLM Swaps, Scales, and Families

Abstract

Most systems that give a language model long-term memory treat memory as a tool the model uses. We invert this: the model is an interchangeable process that reads from and writes to a deterministic memory substrate, and the substrate — not the model — is the durable object. This yields a falsifiable question: when the model is swapped, does the system's account of a user survive? We build Fireweed, a substrate governed by one rule — the model proposes, deterministic code decides, nothing ungrounded is committed — making the memory auditable and byte-reproducible, plus an apparatus that localizes where continuity lives. One might expect the account to drift, since different models carry different extraction biases and world knowledge. Across a 4× parameter-scale gap and two model families, it does not. Different perceiver models build substrates that agree on the canonical entities and on roughly 0.8 semantic claim content (one sentence encoder throughout), against a 0.29 cross-persona floor and a 0.92 same-family ceiling — a cross-family swap costs little more agreement than staying within one family. On a third-party corpus (LoCoMo) the semantic agreement replicates (0.79, +0.43 over a cross-person floor) while discrete entity agreement degrades on open-domain dialogue, localizing which part of the substrate is model-robust versus input-sensitive. A lesion ablating only the substrate's consolidation collapses entity agreement (0.78 to 0.10) while raw claim similarity is unchanged — the structured account is preserved by the substrate, not the shared input. Different reader models paraphrase but agree in meaning (matched similarity 0.7–0.8 versus a 0.33 shuffled-pair floor, p < 0.001). We turn five personal-identity thought experiments — fission, fusion, amnesia, gradual change, transplant — into measurements with null controls, and the account survives all five. Two supporting properties round out the picture: a self-improvement loop whose gains accumulate in the fabric rather than the weights, and structural abstention. A 12-question pilot showed a fabrication reduction directionally consistent across five readers but not significant under cluster-robust resampling; scaled to 1,200 adversarial items over 722 third-party personas, the same pipeline produced zero confident false assertions across 2,400 answer opportunities, against 154 for a bare RAG baseline with the same two readers.

Keywords