Waving Cat Learning Systems
We build memory infrastructure for AI agents.
Agents retrieve by meaning, with no language model anywhere in the retrieval loop. Items are stored as one bit per dimension (at 96 bytes each) and searched by Hamming distance over a 16-million-cell binary lattice. The footprint is fixed: it does not grow with your corpus, and writing new items never triggers a reindex.
One bit per dimension retains 97% of full-precision retrieval quality at 32× the compression, with no training and no learned hash function. That number is ours, it is measured against exact float search, and it is reproducible against any encoder in an afternoon.
Running in daily production: brain cartridges served over MCP, private libraries in the browser, and episodic chains an agent can walk forward and backward.