Major Features #

• HiRAG Phase 2: Complete implementation of advanced hierarchical RAG capabilities.

  • Added LLMAdapter interface for LLM integration.
  • Implemented autoSummarizeLayer() for automatic layer summarization using LLMs.
  • Implemented multiHopSearch() for hierarchical context-aware retrieval.
  • Support for configurable prompt templates and search depth.

• Cross-Device Sync Backends: Real-time synchronization infrastructure.

  • Added SyncBackend interface with pluggable backend architecture.
  • Implemented FirebaseSyncBackend for Firebase Realtime Database.
  • Implemented WebSocketSyncBackend for custom WebSocket servers.
  • Added CrossDeviceSyncManager for managing sync lifecycle.
  • Factory pattern for automatic backend selection.
  • Stream-based real-time change propagation.

• Advanced Re-ranking Strategies: Improve search relevance with multiple algorithms.

  • Added ReRankingStrategy interface for extensible re-ranking.
  • Implemented BM25ReRanker for term frequency-based ranking.
  • Implemented MMRReRanker for Maximal Marginal Relevance (diversity + relevance).
  • Implemented DiversityReRanker for maximizing result variety.
  • Implemented RecencyReRanker for time-based relevance.
  • Integrated re-ranking into semanticSearchWithReRanking() and hybridSearchWithReRanking().

Improvements #

• API Enhancements:

  • Extended MemoryGraph with semanticSearchWithReRanking() method.
  • Extended MemoryGraph with hybridSearchWithReRanking() method.
  • Added createdAt as optional parameter in MemoryNode constructor.
  • Improved query handling in semanticSearch() for layer filtering.

• Architecture:

  • All new classes properly exported in isar_agent_memory.dart.
  • Clean separation between interfaces and implementations.
  • Factory patterns for extensibility.

• Testing:

  • Added hirag_phase2_integration_test.dart for LLM-based summarization.
  • Added multi_hop_retrieval_test.dart for hierarchical search.
  • Added reranking_strategies_test.dart for all re-ranking algorithms.
  • Added advanced_retrieval_test.dart for combined features.
  • Added cross_device_sync_firebase_test.dart for Firebase backend.
  • Added cross_device_sync_websocket_test.dart for WebSocket backend.
  • Added sync_conflict_resolution_test.dart for LWW conflict handling.
  • All core tests passing (13/13 ✅).

Dependencies #

• Added firebase_core and firebase_database for Firebase backend.
• Added web_socket_channel for WebSocket backend.
• Updated mockito to ^5.4.4 for improved testing.

Documentation #

• Updated README.md with comprehensive documentation for all new features.
• Added usage examples for HiRAG Phase 2, sync backends, and re-ranking.
• Updated TASKS.md with completed sprint and new priorities.
• Documented LLMAdapter interface with implementation examples.

Breaking Changes #

None. All changes are backward compatible with v0.3.0.

Notes #

• WebSocket tests require mock generation via build_runner.
• Firebase backend requires valid Firebase configuration.
• LLM adapter implementations require external API keys (e.g., Gemini).
• All sync operations use AES-256-GCM encryption from v0.3.0.

Major Features #

• Sync Protocol: Implemented secure synchronization with Last-Write-Wins (LWW) conflict resolution using UUIDs.

  • Added SyncManager class for encrypted export/import of memory graphs.
  • Added EncryptionService using AES-256-GCM for client-side encryption.
  • Extended MemoryNode and MemoryEdge with sync fields: uuid, modifiedAt, version, deviceId, isDeleted.
  • Documentation: Added doc/SYNC_PROTOCOL.md describing architecture and reconciliation strategy.

• HiRAG (Hierarchical RAG): Foundation for hierarchical knowledge management.

  • Added HierarchicalMemoryGraph extension with createSummaryNode and getNodesByLayer methods.
  • Added layer field to MemoryNode for multi-level knowledge organization.
  • Support for summary and part-of relationships between nodes.

Improvements #

• Model Updates: Regenerated Isar and ObjectBox models to support new sync and HiRAG fields.
• Testing: Added integration test for sync functionality (isar_agent_memory_tests/test/sync_integration_test.dart).
• Testing: Added unit tests for EncryptionService and SyncManager.

Dependencies #

• Added cryptography: ^2.9.0 for encryption support.
• Added json_annotation: ^4.9.0 for JSON serialization.

Documentation #

• Updated TASKS.md to mark release 0.2.3 as completed.
• Added comprehensive sync protocol documentation.

• Fix: Critical bug in OnDeviceEmbeddingsAdapter - corrected outputs.values.first to outputs.first.
• Fix: Corrected invalid Mermaid diagram syntax in README (ASCII art → valid Mermaid).
• Fix: Moved http from dev_dependencies to dependencies for tool scripts.
• Improvement: Replaced forEach with for loops in resource cleanup (better Dart practice).
• CI/CD: Disabled automatic benchmark workflow trigger (manual only) and added Linux desktop support.
• Chore: Added test_resources/ and tool/ to .pubignore to reduce package size.

• New Feature: Added OnDeviceEmbeddingsAdapter using onnxruntime for privacy-first, local embedding generation.
• New Feature: Implemented a basic WordPieceTokenizer for BERT-based models.
• Improvement: Enhanced MemoryGraph robustness with a fallback mechanism (linear scan) when vector index operations fail (e.g., dimension mismatch).
• Improvement: Added explicit dimension validation in ObjectBoxVectorIndex to prevent native crashes.
• Documentation: Translated README.md and TASKS.md to 100% English. Added instructions for On-Device Embeddings.
• Dependencies: Updated langchain to ^0.8.0, objectbox to ^5.0.0, langchain_google to ^0.7.0, and added onnxruntime ^1.4.1.
• Refactor: Removed deprecated vector_index_dvdb.dart.

• CI/CD: Added publish-to-pub-dev.yml workflow for automated publishing on release creation or manual dispatch.
• Credentials: Documented usage of PUB_CREDENTIALS_JSON secret.
• Maintenance: Minor release preparation and local validation.

• Major Cleanup: Removed DVDB as a vector backend.
  • Deleted dvdb dependency and public export.
  • Retained vector_index_dvdb.dart as a deprecated stub (now removed in 0.2.2).
  • Established ObjectBox as the sole supported on-device ANN (HNSW) backend.
• Testing:
  • Introduced InMemoryVectorIndex for plugin-free unit testing.
  • Removed isar_flutter_libs from the test subproject.
  • Fixed similarity metric consistency (cosine/L2/dot) in memory index.
• Documentation:
  • Cleaned up references to DVDB in docs.
  • Clarified ObjectBox default usage.
  • Fixed linting issues.

• Fix: Resolved JavaScript error in Isar generated files by integrating build_runner.

• Documentation: Added comprehensive dartdoc comments to public APIs.
• Linting: Fixed various linting and formatting issues.
• Publishing: Corrected pub.flutter-io.cn topics for successful publication.

• Initial release: Isar agent memory graph with ANN search, explainability, robust tests, and modern CI/CD automation (Coderabbit, Renovate, Dependabot, Jules).