MemoryBox 2.0 Architecture: Zero Built-In Models & 92.6% LongMemEval Long-Term Memory OS
Built in native Rust with embedded_models=false ($0 API tax), MemoryBox 2.0 reaches 463/500 (92.6%) on LongMemEval, 3.64 ms cold start, and model-agnostic portability.
Author: Dalian Fidwell MemoryBox Engineering | Dalian Fidwell Technology Co., Ltd.
Escaping the Vector Lock-In & Embedding Tax Trap
Feeding ~115K raw transcript tokens into frontier models yields roughly 60.2%–60.6% on LongMemEval while burning massive prompt spend per turn. Dense Vector RAG alternatives introduce subtle enterprise failures: confusing 500 ms with 0.5 s or version v2.0.3 with v2.0.4, mixing deprecated policies with current truth, and trapping 10-year organizational memory inside ephemeral model embedding dimensions.
Zero-Model Host Symbiosis & 8+2 Cognitive Lanes
MemoryBox 2.0 (v2.0.4) cleanly decouples reasoning from sovereign storage. Using 8+2 deterministic channels—covering trilingual lexical signals, host concept hints, Kuzu multi-hop hypergraphs, bi-temporal intervals, exact numeric dimensions, speaker roles, version supersession, and macro scene nodes—the engine constructs chronological source bundles in under 30 ms while signaling strict completeness watermarks on bounded recalls.
Verified Real-Hardware Metrics & 3 SKUs
Cold store creation runs in 3.64 ms (0.32 ms warm reopen), ARM64 RK3588 static binaries measure 3.55 MiB with 2.25–2.98 MB idle RAM and 75 µs light recall, and the offline Obsidian-Cosmos WebGL2 console renders 5,000-node graphs in 3.85–5.18 ms/step inside a 57.7 KB zero-CDN bundle.
Related Questions & Answers
How does MemoryBox 2.0 reach 92.6% on LongMemEval with zero embedded neural weights?
Its native Rust kernel unifies deterministic EN/ZH/JA tokenization, an embedded Kuzu cognitive hypergraph, bi-temporal version lineage, exact decimal/unit normalization, and regex-free grounded reading across 8+2 orthogonal lanes—optionally accepting host-supplied synonyms at zero extra API cost.
Will upgrading models or switching tools break existing MemoryBox archives?
No. Unlike 1536-dim vector databases tied to a specific embedding model, MemoryBox indexes are permanent, deterministic structures shared across Claude Code, Cursor, Codex, Trae, OpenCode, and OpenClaw with zero re-indexing.
How does Dream consolidation prevent LLM hallucinations from polluting future recall?
Every Micro-Dream or Deep-Dream summary and relation edge passes a fail-closed persistence gate verifying a non-empty database source span pointer, with one-click atomic batch rollback that leaves raw history untouched.