A test of an AI memory system found a sharp split: entity-specific adapters could store single-answer facts inside a language model, but the system could not recover those facts when it had to choose an adapter from the question. When the correct adapter was supplied directly, answers improved. Two query-driven similarity methods produced no gain over the unadapted model. The result points to a gap between putting knowledge into model weights and finding it later.
The preprint asks whether subgraphs tied to individual entities can be stored without prompt context, whether a query can retrieve the right adapter, and whether adapters can be combined for multi-hop questions. Its design verbalizes each entity subgraph into a LoRA adapter, a learned update tied to that entity, then injects a selected adapter into a frozen base model. At question time, the model answers without the subgraph text in context, a closed-book setup.
Storage worked when the right entity was known
The main bank contained 150 clean single-film entities from MetaQA, selected because each film title mapped to exactly one release year and its one-hop subgraph was internally consistent. The experiment used Qwen3.5-2B with a frozen four-bit QLoRA base. Rank-eight LoRA adapters were trained with AdamW, using eight question-and-answer pairs per entity; four more pairs were held out. Results were compared per entity with paired gains, and 95% confidence intervals were estimated from 50,000 bootstrap resamples.
On held-out single-valued relations, the base model scored 0.007 on exact match, a strict measure that counts only fully matching answers. The adapted model scored 0.250, a gain of 0.243 over base. The 95% confidence interval for that gain was 0.174 to 0.319, with p < 0.001. For this narrow class of one-answer relations, the adapter carried information that the frozen model did not reliably produce on its own.
The improvement was tied to the matching entity. The own adapter gained 0.283 EM over base, with a 95% confidence interval from 0.183 to 0.400 and p < 0.001. An adapter trained for the wrong entity produced a gain of 0.000, while random weights reduced the score by 0.017. The wrong-adapter contrast had p = 1.000, and the analysis cautions that this comparison had limited power, so it should not be treated as proof of equivalence.
The hard part was finding the memory
That result depended on knowing which adapter to use. Retrieval was tested with 30 query entities against a gallery of 120 adapters. One method compared question embeddings with cosine distance. The other compared changes in adapter weights using normalized Frobenius distance over the top layers. In both cases, the top-ranked adapter was injected before closed-book exact match was measured. Neither query-driven method improved on base: each had a gain of 0.000, with p = 1.000. The own-adapter oracle gained 0.283, while random baselines also failed to improve the base, with p-values of 0.778 and 0.533.
The weight comparisons did reveal a meaningful signal. Across the clean bank, weight distance correlated with semantic distance, with a Spearman correlation of +0.329 across all layers and +0.352 across the top eight layers. Label-permutation and column-shuffle controls reached no more than an absolute correlation of 0.157 across 1,000 resamples, with p < 0.001. In other words, the weights reflected semantic similarity, but the tested similarity signal did not translate into useful adapter selection.
The result has a sharp boundary
Multi-answer relations exposed a measurement problem. Exact match was 0.000 for both tags and starred actors, yet alternative scores detected valid partial recovery. For tags, any-match was 0.854 and set recall was 0.793. For starred actors, the corresponding scores were 0.500 and 0.389. The study therefore limits its strongest exact-match storage claim to single-valued relations and leaves set-valued parametric memory for future work.
Composition, the third part of the proposal, could not be judged. The merged or jointly contextualized inputs lacked the required intermediate-entity information, so every tested condition scored at the floor, including an oracle given both subgraphs. No composition effect can be estimated from that experiment. The result is not evidence that multi-hop adapter composition is impossible.
A trade-off, not a deployment verdict
The clearest practical trade-off was in the input and storage accounting. Graph RAG was reported at 60 context tokens and 188 bytes per query. Parametric memory used 0 context tokens at query time but required a 10.4 MiB adapter as a one-time cost. Those figures do not amount to a full deployment comparison: the experiment did not measure end-to-end latency or energy, and adapter injection or swapping may offset the saving in prompt space.
The evidence is narrow. It comes from a single MetaQA movie benchmark, one Qwen3.5-2B model, clean film entities and one-hop subgraphs. Retrieval used only 30 query entities and 120 gallery adapters. These limits mean the results do not establish that the selection methods would fail across other models or domains, or that the storage result would generalize beyond this setting.
What remains to be tested
The next challenge is selecting the right local adapter from a query and combining adapters when a question spans entities. Those are the two steps this experiment did not demonstrate. Until broader tests address them, the work shows a storage result, not a complete memory system.
The work is a preprint presented for the SKGi 2026 workshop.
Paper data and sources
Original title: A Storage-Retrieval Gap in Parametric Knowledge Graph Memory
Authors: Martino M. L. Pulici, Cuong Xuan Chu, Evgeny Kharlamov, Volker Tresp
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text