Krusch Context Mcp
Long-term memory for AI coding agents — remembers project rules across sessions at $0 AI cost (MCP server, SQLite/PostgreSQL). 90.26% on Snap LoCoMo.
Local MCP server: five tools to store, retrieve, and retire project decisions. SQLite by default — powering the
$0.00Zero-LLM Raw-RAG + 4-Way RRF memory substrate behind91.17%(547/600@$24.54,Gemini 4 Argon Low— uploadeddolphinbench.aisubmissionUUID c047b063,status: "queued"awaiting leaderboard acceptance; exceeds current70.67%published leaderboard entry in independent600/600evaluation),90.67%(544/600@$30.54,Argon Low + Grounded v0.3— self-run clean-room evaluation with highestRileyslice181/200 = 90.50%and published SHA-256 audit logs;98.33%Global Oracle@28), and57.33%(344/600@$0.00, Wave-6krush420/gemma4-31b-sre-lora+Grounded— highest open-weights score in self-run 28-run evaluation;54.17%[325/600] stock submissionUUID 288671aduploaded & queued ondolphinbench.ai) across28OfficialDolphinBenchruns (16,800traces),90.26%(1,390/1,540,92.66%Oracle@2,95.77%recall,182 ms) onSnap LoCoMo(ACL 2024), and93.00%10M-token IE recall on ICLR 2026BEAM.
📖 Read the Systems & Benchmark Deep Dives: Read Paper #27 — Return on Memory (
91.17%Queueddolphinbench.aiSubmissionUUID c047b063&90.67%Clean-Room Grounded v0.3) (PDF), Paper #26 — What DolphinBench Taught Me About Local AI (Highest Open-Weights Score57.33%Wave-6 /56.67%Stock+Grounded /54.17%Queued StockUUID 288671adGemma 4 31B@$0.00), and Beyond Flat Vector RAG: What Makes krusch-context-mcp Special (docs/WHAT_MAKES_IT_SPECIAL.md).
From the project's README.