This paper presents empirical evidence from a 2, 100-data-point benchmark across three frontier LLMs (Claude Sonnet 4, GPT-4o, Gemini 2. 0 Flash) demonstrating that the AI industry's debate over file format for LLM memory — markdown vs. structured representations — is directed at the wrong problem. In Stage 1 (N=900), identical facts encoded as flat markdown versus structured relational context produced statistically equivalent accuracy (Δ = -0. 004). In Stage 2 (N=1, 200), full-context markdown significantly outperformed vector RAG, GraphRAG, and hybrid retrieval on strategic queries (0. 964 vs. 0. 888–0. 904, p < 0. 004). Analysis reveals this advantage stems from information completeness: retrieval conditions discarded 84–90% of available context. The paper synthesizes these findings with neuroscience research on reconstructive memory, the data. world knowledge graph benchmark, and production evidence from multi-agent systems to argue that the binding constraint for LLM reasoning is not format but information loss during retrieval — and that at production scale, where corpora exceed context windows, retrieval architecture becomes the determining factor. The accompanying repository includes all code, the 50-document test corpus, 50 queries with deterministic scoring rubrics, and complete raw results for independent reproduction. Total cost to replicate: under 30.
John R. Williams (Tue,) studied this question.