SuperLocalMemory V3.3: Biologically-Inspired Agent Memory
TL;DR
A local-first agent memory system that implements human-like memory dynamics: forgetting curves, cross-session consolidation, 7-channel retrieval, and progressive compression. No cloud LLMs required — runs entirely locally.
Key Findings
- Fisher-Rao Quantization-Aware Distance (FRQAD): a new embedding distance metric on the Gaussian statistical manifold that prefers high-precision embeddings over quantized ones. 100% accuracy vs 85.6% for cosine distance.
- Ebbinghaus Adaptive Forgetting: mathematical forgetting curves in local agent memory. Fading memories progressively lose embedding precision (compression increases as memory fades).
- 7-channel retrieval: semantic, keyword, entity graph, temporal, spreading activation, consolidation, and Hopfield-style associative channels — all running locally.
- Cross-session consolidation: mimics hippocampal-neocortical memory transfer between sessions.
- Zero-LLM operation: full local deployment, no cloud dependency.
Why It Matters
Current agent memory is mostly flat vector databases with single-channel retrieval. This system brings biological memory concepts — forgetting, consolidation, multi-channel association — to local agent memory. The FRQAD metric is particularly interesting as a principled way to handle quantized embeddings (relevant when running models locally with reduced precision).