An experiment with per-entity AI memory found a sharp split between storing information and finding it. When the correct module was supplied, the model answered held-out film questions without seeing the source subgraph text. But the two query-based methods tested to choose that module did not lift performance above the base model.
A separate module for each film entity
The method converted each entity’s verbalized knowledge-graph subgraph into its own LoRA adapter, a separate model add-on trained for that entity. The adapter was injected into a base model whose original weights stayed fixed, and the model then generated an answer without the subgraph text.
The main bank contained 150 clean single-film entities. Each was represented by an internally consistent one-hop subgraph, meaning the test used immediate links around one film rather than longer chains. The model was Qwen3.5-2B with a frozen 4-bit QLoRA base; eight question-answer pairs per entity were used for training and four were held out.
The reported held-out differences were paired against the base model and tested with 50,000 bootstrap resamples, with 95% confidence intervals used to show uncertainty around each gain.
The stored facts were selective
On questions with one expected answer, the adapted model’s exact-match score was 0.250, compared with 0.007 for the base model. The difference was +0.243, with a 95% confidence interval from +0.174 to +0.319 and p < 0.001. Across all relations, the scores were 0.137 and 0.010, for a smaller gain of +0.127, with a 95% confidence interval from +0.083 to +0.170 and p < 0.001.
The biggest relation-specific gains were for director questions (+0.274), writer (+0.269), and release year (+0.196). Genre (+0.182) and language (+0.091) also had positive gains, but the reported confidence intervals excluded zero only for director, writer and release year.
Questions whose answers could contain several items were harder to summarize with exact match. For tags, exact match was 0.000, even though the model achieved 0.854 on an any-match measure, which counts an answer if at least one item is right, and 0.793 on set recall, which measures the share of expected items recovered. For starred actors, the corresponding scores were 0.000, 0.500 and 0.389. A zero exact-match score could therefore coexist with getting part of the answer right without recovering the complete set.
The missing piece was routing
That made adapter selection the practical test. For each of 30 query entities, the system ranked a gallery of 120 adapters using three strategies: distance between question embeddings, normalized Frobenius distance between adapter weights, or random selection. It then injected the top-ranked adapter and measured closed-book exact match.
Neither similarity-based route selected an adapter that raised the score. Question-embedding retrieval and weight-distance retrieval both scored 0.017, for a gain of 0.000 and p = 1.000. Global random selection scored 0.010, with a gain of -0.007; random selection within the domain scored 0.007, with a gain of -0.010. By contrast, the own-adapter oracle, which was given the correct adapter, scored 0.300 for a gain of +0.283.
The results added a complication: the weights did contain some semantic signal. On the clean bank, the ordering of weight-space distances tracked the ordering of subgraph semantic distances with a Spearman correlation of +0.329 across all layers and +0.352 over the top eight layers. Label-permutation and column-shuffle controls produced +0.157 and +0.036. That association carried information, but it was not enough to produce functional retrieval in this test.
A tradeoff with no shortcut yet
The study also tested whether separate adapters could be combined for multi-hop questions. It could not isolate a composition effect because the intermediate entity needed to answer was absent from both the adapter and the context inputs. Every tested condition, including an oracle with both subgraphs in context, scored at floor. The result is a non-result, not evidence that adapter composition works.
The storage tradeoff ran in the opposite direction from the context tradeoff. The parametric route used 0 context tokens per query, compared with 60 for graph RAG, but it carried a one-time adapter cost of 10.4 MiB. Graph RAG shipped 188 bytes per query, while the adapter-to-triple ratio was roughly 58,000. At a bank of one million entities, the estimated adapter checkpoint would be about 10 TiB.
The findings are therefore strongest as a test of this particular arrangement: 150 clean single-film entities, the Qwen3.5-2B base configuration, and a retrieval gallery of 120 adapters for 30 query entities. They do not settle whether the same routing problem would appear in broader settings, and the multi-hop test did not establish successful composition.
The front matter places the work in the 2026 Scaling Knowledge Graphs for Industry Workshop. Its practical question remains open: how to select and combine local adapters from a query when the correct answer is not already known.
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