Metric-audit data for S-VEAM: A Shared Spectral Memory for Cross-Session Retrieval in Long-Horizon LLM Agents (CIKM 2026)

Metric-audit data accompanying the CIKM 2026 paper "S-VEAM: A Shared Spectral Memory for Cross-Session Retrieval in Long-Horizon LLM Agents" (DOI 10.1145/3799682.3840582). For the cells behind the reported tables the package holds the per-method, per-query retrieval records, the aggregated metric files, and the run configuration — every argument, together with the commit each run was built from — so each table in the paper can be inspected and re-aggregated without rerunning any retriever. 125 cells, 179,878 per-query rows. Data only; no code. The configuration that produced the numbers is stated in full in the paper's method section. Benchmark text is omitted. The retrieved context, the questions and the gold answers belong to LoCoMo, LongMemEval-S and MuSiQue, and redistributing them is not ours to grant; LongMemEval-S is additionally gated at its source. What remains is sufficient to re-derive every reported number: per-query metrics, the retrieved and gold-evidence session ids, and instance_id, which joins each row back to its source instance. One scope note. Table 3's three architectural rows — indexing, lexical-SVD and persistence removal — were produced by an earlier collation whose per-query records are not in this package. The directory table3_axes_rerun holds a later, complete run of the same three axes (n=1542, with per-query records); its deltas reproduce the printed rows to within 0.0026, inside the paper's 0.005 reporting tolerance, but it is not the run the printed values came from and is labelled accordingly. MANIFEST.json indexes every cell with a sha256 per file.