🚀 Benchmark Spotlight: Spector Memory Achieves 83.42% Accuracy & 12ms Recall Latency on LoCoMo

[sbharatjoshi]

🚀 Spector Memory: Benchmark Results on LoCoMo

We are excited to share the latest downstream generative QA evaluation results for Spector Memory on the LoCoMo (Long-Context Memory for Multi-Turn Dialogues) benchmark.

Spector Memory achieved an 83.42% Overall J-Score (Accuracy) while delivering an ultra-compact context injection and ~12ms sub-system memory recall latency directly from its native off-heap memory substrate.

---

📊 Headline Performance & Baseline Comparison

Metric	Spector Memory (Observed)	Zep (Published)	Mem0 (Published)
Overall Accuracy (J-Score)	83.42%	75.14% – 80.00%	62.47% – 68.20%
Context Tokens Added	1,257 tokens	3,911 tokens	1,764 tokens
Pure Memory Recall Latency	12.06 ms (Min: 9.88 ms)	632.0 ms	657.0 ms

Key Takeaway: Spector Memory achieves higher recall accuracy while injecting ~3.1× fewer context tokens and running >50× faster memory retrieval than traditional graph/vector-only architectures.

---

🧠 Cognitive Subsystem Reasoning Breakdown

Spector's biological multi-layer architecture routes and activates different cognitive subsystems depending on the reasoning demands of the query:

Reasoning Subsystem	J-Score (Accuracy %)	Context Tokens	Pure Memory Latency
🏷️ Entity Graph & Directory	91.43%	1,255 tokens	~11 ms
⏳ Temporal Knowledge Graph & Chains	86.49%	1,283 tokens	~10 ms
🕸️ HyperEntity Multi-Relational Graph	76.92%	1,271 tokens	~11 ms
🧠 Hebbian Associative Spreading	75.95%	1,245 tokens	~12 ms

---

⚡ Architectural Highlights

1. Zero Pre-Computed Offline Graphs: Data is ingested 100% naturally through the runtime memory.remember() pathway. Synaptic tagging, dynamic entity directory indexing, hyperedge construction, and temporal knowledge graph facts are generated autonomously on the fly.
2. Native Java 25 Foreign Function & Memory (FFM) Substrate: Off-heap direct slab layout (HeaderLayout64) + AVX-512 SIMD enables sub-10ms warm search across 500+ dialogue turns without garbage collection overhead.
3. Information-Theoretic Attenuation & Lateral Inhibition: ACT-R fan-factor spreading activation ensures multi-hop knowledge retrieval without flooding context windows with noise.

---

Tested on LoCoMo multi-turn dialogue corpus with Google Gemini 3.1 Flash-Lite.