amanhij/Zikkaron

Implements an MCP server architecture with 24 tools for Claude Code integration, using a 22MB embedding model and SQLite for local persistence without requiring API calls or large language models at query time. The system uses Hopfield energy scoring, spreading activation, thermodynamic decay, and cross-encoder reranking to achieve 86.8% recall on standard benchmarks and 0.404 overall accuracy on the 10M-token BEAM benchmark—a 52% improvement over prior systems. Automatically syncs instructions via `sync_instructions` and provides hippocampal replay consolidation to preserve context nuance during Claude Code's native compaction of long sessions.

Available on PyPI.