Cognitive Audit.
Measured performance audit for the Recallix MemoryOS. 0.106ms retrieval at 1,000,000 nodes. Verified on Apple M4 Core Infrastructure.
Writes
Recall@5 (1k)
500-Turn Retention
Multi-Agent Fidelity
Memory Accuracy
100% Recall@1 and Recall@5 on hardened benchmark. 300 facts stored with semantic distractors per entity. Queries designed without keyword overlap to force genuine semantic retrieval.
"Recallix successfully isolated core episodic events from high-noise conversational buffers, maintaining 20/20 recall with zero distracter leakage."
Measured Search Latency
Average C++ Search (1M Nodes)
0.106ms pure C++ search at 1M nodes. ~1-5ms end-to-end API latency at production scale. Two-phase async architecture ensures writes never block agents. Benchmarked on Apple M4 with NEON SIMD hardware acceleration.
Universal Substrate
The Cross-Model Test
Knowledge was stored using Llama 3.1 8B. Retrieval was performed using Mistral 7B.
100% exact match fidelity across different model architectures. Memory store latency: 407ms API. LLM generation time excluded from memory metrics.
Context Retention
75% retention after 500 conversational turns with semantic collisions and evolving facts.
Intent Shadowing on low-importance date facts. Adaptive semantic fallback in development to resolve episodic matching collisions.
Scale Analysis
Optimized for agent workloads under 10k memories. Improving 100k+ recall actively.
Adversarial Gauntlet
10/12 Adversarial Tests Passing
2 known failures documented and tracked for the next core sprint.