CedroDB 0.6.0 — agent checkpoints and hybrid memory

CedroDB 0.6 adds local memory retrieval and revision-protected checkpoints for application-managed AI agents. Save a checkpoint and ordinary result/outbox documents in one durable transaction, then reopen the same thread after restart. A stale continuation receives a conflict and writes nothing.

• Search JSON memory by text/BM25, supplied vectors/exact cosine, or hybrid/RRF. Metadata prefilters run before scoring; equality indexes narrow candidates; scan-budget exhaustion is an explicit error.
• Use matching native Rust, Node.js/TypeScript and Python APIs with typed results, deterministic ranking ties and separate scoring details.
• Preserve exact floating-point values during replay and restore, fixing possible one-ULP changes in earlier releases.
• Run the pinned PydanticAI 2.56.0 conversation example with TestModel: conversation identity, messages and usage survive reopening, with no model-service calls.
• Explore the recorded Rust agent demo and eight Node/Python recipes on the portfolio.

Validation: 104 Rust, 62 Node and 66 Python tests locally; native SDK builds, interoperability, optional PydanticAI recipe and source distributions are checked in the release CI. SHA-256 sidecars accompany each distribution.

The existing log and backup formats are unchanged. _cedro.agents.* is newly reserved; migrate conflicting ordinary collections with 0.5 before upgrading. This release remains experimental. Retrieval scans committed in-memory documents; no ANN index, model inference, framework checkpointer protocol, tool authorization or automatic workflow resumption is provided.

Agent contracts · Research and positioning · Migration notes