Changelog.

The local and mcp extras now install sentence-transformers by default, preventing clean installations from silently using deterministic test-grade embeddings.

Regression tests now require semantic embeddings in the local, mcp, and all extras.

Python, TypeScript, and Java package pages now lead with bitemporal recall, deterministic supersession, audit history, and runnable evaluation evidence.

The integration now uses the official agents import path and documents all four memory tools.

The preprint repository now includes a versioned JSON Schema and a validated five-invariant Lians live run.

Durable facts persist into governed memory before the host framework compacts context, automatic via CompactionGuard, with flush nodes/tools for LangGraph, the OpenAI Agents SDK, CrewAI, and AutoGen.

An agent's point-in-time knowledge state as a human-readable document, its SHA-256 anchored in the audit chain, so the statement you hand an examiner is itself tamper-evident.

Unresolved conflicts ride at the top of every assembled context block as explicit "X disagrees with Y" lines.

Embedding outages no longer take recall down: queries proceed lexical-only, with the degradation declared on every result and audit event.

Retrieval half-life scales with materiality: passing preferences fade in 7 days, critical flags hold for 365.