A RAG system can have the right fact in memory and still not answer. The issue is not the memory itself, but the artifact the model receives as input: a memory entry, a summary, a typed record, or a raw excerpt. RENDER, the control introduced by the researchers, fixes the conversation and varies only this layer. Across 500 LongMemEval questions and nine models, matched-budget resolved packets beat recency-truncated raw dialogue by 42.4–72.6 points. In deployed-style templates, the best-worst spread ranges from 24.6 to 48.8 points per model. Under the primary scorer, ChatGPT-style entries have higher point estimates than raw conversation on seven of nine models. Judge rescoring preserves the positive aggregate effect, but model-specific significance is mixed.

The most revealing result is another one: three models that score zero on formal ledger packets answer the same facts when they are expressed in natural language in 45.4–53.4 percent of cases. This is not a knowledge problem, but a rendering problem. The same content becomes usable only when it arrives in a form the model internalized during training.

The artifact is not an implementation detail

Many memory and RAG benchmarks treat the model's input as an internal detail. RENDER shows it is a variable with effects in the same order of magnitude as other evaluation choices. A five-level scale localizes where answer-bearing content enters the input; deterministic templates approximate ChatGPT-style entries, LangChain-style summaries, MemGPT-style typed records, and raw dialogue. The variation is not noise: it is systematic, persists under retrieval noise, and transfers to HotpotQA.

For teams running self-hosted stacks, this has a concrete consequence. A structured artifact can look more efficient because it is compact and auditable, but if the model cannot read it, the efficiency is only apparent. At the same time, a longer natural-language entry consumes more input tokens, with direct effects on VRAM, throughput, and latency. The point is not which format is better in absolute terms, but that the decision cannot be separated from evaluation. AI-RADAR provides analytical frameworks on /llm-onpremise for assessing these trade-offs.

The structural implication is that benchmarks should report or control the reader-facing artifact. Otherwise rankings remain unstable: a model evaluated on raw dialogue can appear weak on structured records and vice versa. This shifts the burden onto teams building memory layers for on-premise deployments: without a per-artifact matrix, model choice depends on a hidden assumption about the format the system will use in production. It is not an academic question: it is the point where evaluation stops describing the system and starts describing only the benchmark.