Agent-memory products expose different data, retrieval, and operations contracts. SuperLocalMemory combines a local core data path with public research preprints and release-linked verification gates.
Competitor behavior, pricing, licenses, and benchmarks change. This page defines the same verification questions for every system instead of publishing an unversioned winner table.
| Dimension | Current SLM contract | Proof required for any product |
|---|---|---|
| Retrieval | Five candidate producers, fusion, optional reranking, graph and Fisher-informed enhancement | Frozen corpus, channel traces, relevance labels, and raw outputs |
| Score meaning | Score Contract v2; answer confidence is null while calibration is pending | Field definitions, held-out calibration, selective-risk, ECE, and Brier reports |
| Data path | Mode A local core; optional providers, connectors, backup, proxies, and downloads are separate | Complete configured flow, retention, erasure, logs, and provider terms |
| Cache | Exact cache opt-in; semantic cache experimental | False-hit, invalidation, tenant-isolation, latency, and provider-accounting evidence |
| Compression | Safe JSON/code paths may yield no reduction; aggressive prose is opt-in and lossy | Declared corpus, tokenizer, ratio, round trip, quality, and latency artifacts |
| Scale | No current V4 capacity or constant-latency claim | Hardware, corpus, concurrency, ingest, p50/p95/p99, quality, disk, RAM, and recovery |
| Client support | Named configurations; verified count not yet published | Install, session, remember, recall, update, forget, restart, upgrade, and uninstall |
Current competitor comparisons must be regenerated from live primary sources and release-linked artifacts before publication. The historical SLM methodology remains in arXiv:2603.14588.
Public preprints, source code, and a V4 ledger make the evidence boundary inspectable.
Dense candidate generation uses cosine similarity. Later fusion, reranking, graph enhancement, and Fisher-informed terms are separately visible in runtime diagnostics when available.
Mode A can run core memory operations without a cloud model-provider call. Optional providers, connectors, backup, proxies, model downloads, and dependency acquisition have separate network behavior.
Three public arXiv preprints document versioned experiments. They are not venue-reviewed, and historical results are not current V4 measurements.