EMPIRICAL BENCHMARKS & REPRODUCIBILITY

Verifiable Systems Benchmarks across 10,000 Notes

Sphene is engineered for deterministic, sub-millisecond local-first agent workflows under 50MB RAM (<25MB local-only). Below are empirical benchmarks measured on standard Linux x86_64 hardware comparing native systems engineering against heavy client runtimes.

Summary Benchmark: Sphene delivers <0.20ms SQLite FTS5 search queries, <1.5ms direct Stdio MCP tool invocations, and zero lock contention under concurrent agent writes, while maintaining an active footprint of under 50MB RAM (<25MB local-only).
Performance Metric Heavy Desktop / Shell-Wrapper Paradigm Sphene v2.2 Native Systems Engine
Daemon RAM at Rest 400 MB – 1.2 GB (Electron / Chromium web view) < 25 MB RAM (Local daemon; <50MB with WireGuard relay)
Peak Indexing RAM (10k notes) 1.4 GB – 2.6 GB (High GC churn & DOM tree) < 25 MB RAM (Bounded batch indexing)
Full-Text Search Latency 45 ms – 180 ms (Unindexed file / regex scan) < 0.20 ms (Embedded SQLite FTS5 WAL)
Intra-Document AST Extraction Full document loaded into context < 0.05 ms (Pinpoint heading subtree)
Agent Tool Invocation Roundtrip 120 ms – 380 ms (Bash subshell / pipe / curl) < 1.5 ms (Direct Stdio JSON-RPC MCP)
Concurrency Failure Rate Frequent file lock & race exceptions 0.0% (Zero-Lock WAL) (5s busy queue + pool)
Agent Token Window Waste 100% whole-file dump per update Up to 95% reduction (AST-level section patch)
Cold-Boot Startup Time 2,500 ms – 6,000 ms (Framework initialization) < 15 ms (Instant native Go binary)
🔬

Testing Methodology & Hardware Baseline

All benchmarks are executed on commodity hardware without synthetic advantages:

Hardware Platform

Intel Core i7-11800H @ 2.30GHz, 32GB RAM, standard NVMe SSD storage, running Ubuntu 24.04 LTS.

Dataset Characteristics

10,000 markdown notes containing YAML frontmatter, markdown headings, code blocks, task lists, and bidirectional wikilinks (~142MB raw text).

Reproducibility Commands

Verified using sphene index, sphene search, and standard Go benchmarking suites (go test -bench=.).

Experience Sub-Millisecond Knowledge Retrieval

Deploy Sphene locally in 60 seconds with native MCP tools enabled.

Get Sphene Free (v2.2)