The problem
AI systems need durable, trustworthy knowledge. Plain fact storage throws away how something was known, whether it's still valid, or what to do when two sources disagree — most memory layers let you silently overwrite or delete history.
The core layer
KGK supplies application-neutral laws for durable knowledge: append-only events, immutable statements, provenance, and soft retraction that invalidates without erasing. It also tracks bi-temporal views (real-world validity time versus system observation time), epistemic status, conflict handling, entity consolidation, and hybrid retrieval. Applications decide domain meaning and authority tiers; the kernel just guarantees how assertions entered and changed state.
The implementation
It's a Python kernel exposed via a Python API, CLI, and FastAPI, with exchangeable in-memory and SQLite backends behind narrow interfaces. Retrieval combines lexical and vector fusion with graph traversal. It's verified with pytest and ruff, and demonstrable end to end: ingest two conflicting assertions, inspect the conflict view, retract or supersede one, then replay the event ledger to confirm it reproduces the same projection.
0 likes · 16 views · Aug 25, 2026
A modular, traceable knowledge graph kernel for AI agent reasoning.