Published V3 LoCoMo Evidence Carried into V3.7

This page is the canonical public reference for SuperLocalMemory benchmark figures. It reports the published V3 architecture experiments that remain the foundation of V3.7. These are not a newly rerun V3.7 package benchmark and are not a cross-vendor leaderboard.

Source: SuperLocalMemory V3 preprint (arXiv:2603.14588).

Published LoCoMo results

Published configuration	LoCoMo aggregate	Original protocol scope	What the result establishes
Mode A Raw	60.4%	10 conversations; 1,276 scored questions; local embeddings, local retrieval, and zero-LLM answer construction	End-to-end local answer construction under the published V3 protocol.
Mode A Retrieval	74.8%	10 conversations; 1,276 scored questions; local retrieval, then GPT-4.1-mini answer synthesis	Retrieval evidence from the local system, with the disclosed external model constructing the final answer.
Mode C	87.7%	Conv-30 only; 81 scored questions; text-embedding-3-large plus GPT-4.1-mini answer generation and judge	Cloud-assisted configuration on one disclosed conversation; not a full-dataset result.

The Mode A Retrieval category results were 72.0% single-hop, 70.3% multi-hop, 80.0% temporal, and 85.0% open-domain. Mode C does not report a temporal category in this one-conversation scope.

Mathematical ablation evidence

Across six LoCoMo conversations, the paper reports 71.7% with the information-geometric layers and 58.9% without them: +12.7pp. This is an ablation result with its own protocol; it must not be substituted for the 10-conversation aggregate rows above.

How to compare responsibly

A LoCoMo percentage is comparable only when the conversation subset, question count, retrieval stack, answer-construction model, judge, and release artifact are declared and compatible. Do not use these figures to claim that V3.7 has been freshly rerun, that every mode has the same score, or that SLM outranks a competitor tested under another protocol.