RC-1 — Neural Memory v0.1.0-rc1
Production-grade semantic memory system for Hermes Agent.
What's in This RC
Architecture
C++ SIMD ──ctypes──▶ Python orchestration
│ │
▼ ▼
MSSQL (primary) SQLite (fallback)
GraphNodes/Edges Dream Engine sandbox
AVX2 similarity Session tracking
SpreadingActivation
Core System
- Semantic recall via
BAAI/bge-m3(1024d), CUDA-accelerated - Knowledge graph with automatic connection + conflict detection
- Spreading activation for idea exploration (BFS + decay)
- Dream Engine — 3-phase autonomous consolidation (
NREM → REM → Insight)
Performance Layer
- C++ SIMD bridge — recall under 1ms
- Cython
fast_ops— 66×cosine_similarityspeedup - LSTM+kNN pattern learning — C++ backends; learns temporal access patterns,
re-ranks recall by embedding + temporal + frequency + graph signals
Storage
- C++ owns MSSQL exclusively — Python never touches
pyodbcdirectly - SQLite fallback for lite mode
- Auto-detects backend via
config.yaml(single source of truth) - Schema:
NeuralMemory+GraphNodes+GraphEdges
Benchmarks
Harness: EvoMem + LoCoMo · local llama-server · no rate limits
Category-specific prompt engineering: temporal / adversarial / general
| Suite | Δ with Dream Engine |
|---|---|
| MMLU | +25.0% |
| GPQA | +46.7% |
Modes
| Mode | RAM | Backend | GPU |
|---|---|---|---|
| Lite | ~50 MB | SQLite + hash/tfidf | ✗ |
| Full Stack | ~500 MB | MSSQL + bge-m3 + C++ LSTM+kNN | ✓ |
Tools (6)
neural_remember neural_recall neural_think
neural_graph neural_dream neural_dream_stats
Known Issues
| # | Issue | Note |
|---|---|---|
| 1 | Prefetch disabled | Returns unfiltered garbage — needs session-aware filtering |
| 2 | C++ SIMD + PyTorch libgomp conflict → segfault
| Use use_cpp=False for benchmarks
|
| 3 | 3-copy file sync required | python/ ↔ hermes-plugin/ ↔ ~/.hermes/plugins/
|
What This Isn't
Not a demo. Not a prototype.
70 commits · 38K lines · 26 bugs found the hard way.
Every fix has scar tissue and a skill to prove it.