It acts as a Git-like layer for Python agents: unlike standard vector memory, it enforces Pydantic schemas and supports transactional rollbacks (Time Travel) to prevent state corruption.
Key technical features:
* Strict Schemas: Uses Pydantic to validate memory on write. If an LLM tries to save a string into an int field, the transaction fails before corruption happens.
* Time Travel: Every change is an append-only transaction. You can call memory.rollback(1) to undo an agent's hallucination.
* Constraints: Enforce logic like Singleton facts (e.g., only one "User Profile" per session).
* No Vectors required: It uses SQLite's JSON1 extension for structured filtering, which is faster and deterministic for state lookups.
It includes a native checkpointer for LangGraph, so you can persist agent threads to SQLite with full history auditability.
It's open source (Apache 2.0). Would love to hear your feedback on the architecture!