Skip to content

CHANGELOG.md

Latest commit

 

History

History
131 lines (104 loc) · 6.01 KB

CHANGELOG.md

File metadata and controls

131 lines (104 loc) · 6.01 KB
Raw

Changelog

All notable changes to the Memori Python SDK will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[Unreleased]

Added

  • Added input validation for the query argument of Memori.recall(...): a non-string query now raises TypeError and an empty or whitespace-only query raises ValueError, matching the existing limit validation and failing fast instead of issuing an empty recall against the database/LLM path.

3.3.6 - 2026-05-27

Added

  • Added MCP client setup guidance for project-scoped attribution using workspace-derived values for X-Memori-Entity-Id and X-Memori-Process-Id to prevent memory mixing across projects. (Refs #404)
  • Added TiDB Zero BYODB provisioning via Memori.provision(...), the python -m memori provision CLI command, and the tidb-zero optional dependency extra.

Changed

  • Local embeddings now use the native Rust fastembed backend exclusively, including Python Memori.embed_texts(...), recall query embeddings, and advanced augmentation fact embeddings. The Python sentence-transformers fallback has been removed; the embeddings optional extra remains as a no-op for install compatibility.
  • Advanced augmentation fact embeddings are attached in the Rust worker before persistence, and the Python write callback reuses the active rust-core engine when available.

Fixed

  • Advanced augmentation fact writes no longer depend on Python sentence-transformers or its transitive dependencies.
  • Rust augmentation now logs when fact embedding attachment is skipped due to a row-count mismatch.
  • Rust-backed BYODB recall now serializes nested recalled summaries before passing rows into the native engine, preventing TiDB Zero/MySQL datetime values from raising TypeError: Object of type datetime is not JSON serializable.

3.3.2 - 2026-04-28

Added

  • Android wheel build coverage for the Rust-backed Python extension, targeting android_24_arm64_v8a and android_24_x86_64 via cibuildwheel. The Rust core ONNX Runtime bootstrap can now download and select the matching libonnxruntime.so from Microsoft's Android AAR at runtime.

Changed

  • Rust-backed retrieval and augmentation are now enabled by default for BYODB mode when memori_python loads successfully. The Python SDK still orchestrates provider wrapping, storage adapters, conversation persistence, and fallbacks. Override with use_rust_core=False on Memori(...), MEMORI_DISABLE_RUST_CORE=1, or legacy MEMORI_USE_RUST_CORE=0; unsuccessful Rust core loads still fall back to the pure-Python path with a warning.

Fixed

  • Conversation injection no longer corrupts OpenAI-compatible message sequences for tool-using conversations. Recalled history previously replayed role="tool" rows (with no tool_call_id) and empty assistant rows (whose tool_calls were never persisted), causing upstream providers to reject the request with 400: An assistant message with 'tool_calls' must be followed by tool messages responding to each 'tool_call_id'. The OpenAI/Anthropic/Bedrock injection paths now sanitise recalled history before prepending it. Legacy Gemini-era role="model" rows are normalised to role="assistant" for the same reason. The injected-message counter now tracks the post-sanitisation count so the post-response payload does not slice into the current user message before persistence and augmentation. (#434)

3.3.0rc1 - 2026-04-16

Added

  • Experimental Rust-backed retrieval and augmentation path. Opt-in via MEMORI_USE_RUST_CORE=1 in BYODB mode. Provides a native hybrid-search recall pipeline (dense + lexical re-ranking) and lower-overhead background augmentation dispatch via a tokio-based worker runtime. The pure-Python path remains the default and is unchanged.
  • Prebuilt platform wheels. Released as cp310-abi3 wheels covering Python 3.10 through 3.14 on manylinux_2_28_{x86_64,aarch64}, macosx_{x86_64,arm64}, and win_amd64. First-time users of the Rust core will download a ~25 MB ONNX embedding model from Hugging Face on first use, cached under ~/.fastembed_cache/.
  • Source distribution now includes the core/ Rust crate. Users on unsupported platforms (or who opt out of the wheel) can build from source provided a Rust toolchain is available.
  • Continuous integration for the Rust core (core-ci.yml) covering cargo fmt, clippy, unit tests, and cross-platform wheel build smoke via cibuildwheel.
  • dry_run and publish_memorisdk inputs on the PyPI publish workflow for release dress rehearsals without touching the index.

Changed

  • Internal crate directory rust-core/ renamed to core/. No public import path is affected; the Rust crate name (engine-orchestrator) and the Python extension name (memori_python) are unchanged.
  • Debug payload logging in the augmentation pipeline now routes through the standard logging module at DEBUG level. The previous stdout-based behavior gated by MEMORI_DEBUG_AA_PAYLOAD=1 has been removed; enable debug-level logging on the memori._rust_core and engine_orchestrator loggers instead.
  • PyPI publish pipeline rewritten around cibuildwheel v3.4.0 for PyPI-compliant wheel tags across all supported platforms. Pure-Python fallback is still available via the sdist.

Fixed

  • Fixed multi-turn conversation ingestion for AzureOpenAI and OpenAI clients. Previously, only the first conversation turn was being recorded. Now conversation_id is resolved early in the request lifecycle, ensuring all conversation turns are properly ingested into the same conversation. (Fixes #83)