Model upgrades are routine; memory migrations are not. A new research paper, still under review, shows that an AI agent can keep the exact same memory store and still effectively forget what it knew once the underlying model changes.
The study, titled "Does Your Agent's Memory Survive a Model Upgrade? A Controlled Study of Memory Portability" and posted to arXiv on 4 September 2026 by Ankit Goyal and Jaideep Ray, compares four ways of preserving the same agent history: verbatim long-context reading (LC-RAW), chunked retrieval-augmented generation (RAG), model-compressed natural-language notes (NOTES), and a fixed-schema knowledge graph (KG-fixed). The work is research-stage: it uses 48 synthetic histories with randomized answer codes, exact scoring, and two open-weight models with fewer than 10 billion parameters.
What the measurements show
The headline result is that structure survives and compression does not. Fixed-schema knowledge graphs transferred almost perfectly: KG-fixed accuracy changed by only +0.0004 ± 0.0020 after a writer-model swap. Compressed notes, by contrast, showed high model coupling — accuracy shifted asymmetrically by +9.91 or −13.28 percentage points depending on which direction the migration ran. The same store, in other words, can gain or lose double-digit accuracy purely because a different model wrote or read the notes.
RAG systems failed differently. A partial embedding migration using a 50/50 mixed index captured only a 4.96-point accuracy improvement, forfeiting most of the 11.90-point gain that a full re-embedding delivered. Mixing old and new embedding spaces, the results suggest, is close to a worst-of-both-worlds option.
The authors also decompose where the damage comes from. About 80% (0.467 ± 0.014) of the NOTES accuracy deficit traces to information lost during the initial compression, while retrieval failures drive 81% (0.364 ± 0.012) of the RAG deficit. Different memory formats break in different places, so a single migration playbook will not cover both.
Repair only works if you kept the raw history
The paper's most operationally pointed finding concerns recovery. Store-only repair of NOTES — trying to fix compressed notes without the original material — failed to reach a 90% performance recovery target in all 48 test cases. When the raw source history was retained, recovery succeeded in 34 of 48 cases for one tested migration direction. Once the original evidence is gone, the authors' measurements indicate, the lost information is largely unrecoverable.
The authors conclude with three recommendations: direction-specific migration testing (because the asymmetry means A→B tells you nothing about B→A), strict embedding space isolation, and retention of source histories to make repair possible at all.
Why it matters
The timing is not academic. Memory has become a first-class layer of the agent stack, with standardized benchmarks such as LoCoMo, LongMemEval, and BEAM now used to compare architectures, and vendors reporting scores like 92.5 on LoCoMo at roughly 6,900 tokens per query. Gartner projects that 40% of enterprise applications will integrate task-specific AI agents by the end of 2026, up from less than 5% in 2025. Every one of those deployments will eventually face a model upgrade — and this study suggests the memory layer, not the model, is where the silent failures will live.
There is also a lock-in dimension. Memory compounds over time and is widely seen as the stickiest form of provider lock-in; if compressed memories are coupled to the model that wrote them, switching models means re-deriving memory from raw history — which only works for teams that kept it.
Unknown
The study is a preprint under review, using 48 synthetic histories and two sub-10B open-weight models. How far the portability gaps generalize to proprietary frontier agents, production memory products, or larger model swaps is not established here. Vendor blog benchmarks cited in secondary coverage are not part of this paper's evidence set.
Our take
The most actionable number here is 0 of 48: store-only repair never hit the recovery target. That effectively turns "keep your raw conversation history" from a nice-to-have into a hard architectural requirement for any long-lived agent. The caveat worth holding: these are sub-10B open-weight models on 48 synthetic histories, so the effect sizes may not transfer directly to frontier production models — but the asymmetry and the repair failure are structural mechanisms, not benchmark quirks, and they deserve replication at scale before anyone migrates a production memory store.
