Astrum Verum – A Vector Symbolic cognitive memory that beats RAG
Astrum Verum is a research project focused on developing a cognitive memory architecture for AI agents. It consists of two phases, with the second phase utilizing Vector Symbolic Architectures to improve memory retrieval and reduce hallucinations. The project aims to enhance the structural recall of facts compared to traditional memory systems.
- ▪Astrum Verum's Phase 1 attempted to organize memory using geometric lattices but was found insufficient for structural recall.
- ▪Phase 2 employs Vector Symbolic Architectures, allowing for associative and structural retrieval of information.
- ▪The new memory engine, CognitiveMemory, can recover facts from corrupted cues and distinguishes between roles in relationships.
Opening excerpt (first ~120 words) tap to expand
Astrum Verum Composition-episodic cognitive memory for AI agents — and an honest record of how it got here. Astrum Verum is a research project containing two distinct phases of memory architecture development. It started as an attempt to organize memory on perfect geometric lattices (Phase 1), but when that proved insufficient for structural recall, it pivoted to Vector Symbolic Architectures (Phase 2). Both phases ship in this repository. Phase 1 is kept as a documented historical mockup. Phase 2 is the working, validated engine that powers the actual AI agent. Read the new mathematical paper: Why VSA Works for Large-Scale Memory (Solving Capacity Collapse & Decoding Hallucinations) Full story, math and results: docs/astrum_verum_design.md.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at GitHub.