Changelog

All notable changes to MemOS will be documented here
GitHubRelease Notes
v2.0.10
Improvements
PolarDB Database:
  • PolarDB optimization and upgrade. Optimized execution paths for queries with multiple filter conditions, leveraging new kernel capabilities to improve CPU and memory usage, achieving several-fold query acceleration and reducing system pressure.
Memory Version Service:
  • Memory version plugin-based architecture.
  • Added high availability for NLI model.
  • Added routing logic for memory version at memreader. Falls back to legacy prompt and model when no relevant memory is recalled or all recalled content is judged irrelevant by the NLI model; otherwise uses the memory version plugin (qwen-flash).
Service Optimization:
  • Optimized scheduling tasks to improve memory usage.
v2.0.9
Improvements
Memory version compatibility for feedback:
  • When Feedback updates a node, it follows the memory version logic, updates the original node, and pushes the previous content into the history field.
Memory version optimization:
  • During conflict resolution and duplicate merging, the process of "restoring independent facts from past versions" has been comprehensively optimized. It can now restore independent facts more stably and accurately, preserve the completeness of memory information, and reduce memory inconsistencies caused by unstable model responses.
Bug Fixes
Memory version parameter and related bug fixes:
  • Fixed several bugs including memory version parameter validation, empty key after update, and incorrect update time.
Feedback issue fixes:
  • Fixed the issue where sources were empty during the Feedback process.
v2.0.8
New Features
New Features:
  • When adding memories, conflicting/duplicate memories are automatically detected, merged, and archived. The latest memory content is retained, while the old memory content is placed in the history field.
Improvements
Preference memories migrated to graph database:
  • Preference memories have been migrated to the graph database, unifying them with other memory types.
memory versioning optimization:
  • Memory retrieval reduces redundant calls to get_by_metadata, decreasing retrieval latency.
database optimization:
  • Implemented fine-grained concurrent rate limiting using semaphores, combined with automatic faulty connection removal and thread lifecycle management, to enhance system stability under high concurrency.
  • Optimized log levels and content, retaining key execution traces while reducing the impact of log writing on database performance.
  • Implemented trigger-based preheating to load hot data into the cache, avoiding high latency caused by cold starts.
Bug Fixes
bug in fast memory addition:
  • Fixed an issue where, when adding memories with MultiModalStruct selected, embedding operations were not performed in some cases under fast mode, causing memories in fast mode to be unsearchable.
bug in multimodal memory:
  • Fixed a bug where the image content was not being separately interpreted for image-based memories.
Memory Retrieval:
  • Reduced redundant calls to get_by_metadata during memory retrieval, decreasing retrieval latency.
Memory Versioning Concurrency Fix:
  • Memory versioning now, under concurrent add scenarios (receiving multiple related add requests within a time frame shorter than the add interface latency), will sequentially merge and process them in chronological order instead of falling back to no processing.
v2.0.7
Improvements
Parameter optimization:
  • The search and chat APIs have the relativity parameter for filtering recalled memories by relevance score. This update sets the default relevance threshold to 0.45.
Log optimization:
  • Intercept some log errors and adjust the log level for certain logs.
v2.0.6
New Features
Retrieval Optimization:
  • Added keyword-based search in the search phase, while relativity continues to use cosine similarity of embeddings.
  • Added a relativity output field to the Chat API, allowing users to control the retrieval threshold during conversations using relativity.
Bug Fixes
Fixes:
  • Fixed an issue where an incorrect field was used in preference memory threshold filtering.
  • Fixed an issue that caused failures or hanging when calling the get_memory API with the filter parameter containing both field conditions and logical operators such as AND/OR/NOT.
  • Fixed issues related to passing extra body parameters to LLMs, ensuring compatibility with the enable thinking parameter input
v2.0.5
New Features
Knowledge Base Retrieval Optimization:
  • Dual-track retrieval: Supports coordinated storage and retrieval of "original text + memory", with customizable recall modes.
  • Context wake-up: Supports recalling surrounding context before/after chunks to improve semantic coherence in long texts.
  • Relevance filtering: The search API adds a relativity field (0~1) to filter low-quality recalls by threshold.
MindDock:
  • Cloud Skill: Added cloud skill support and a quick entry.
  • Real-time prompt injection: Dynamically inject Skill modules during chat to enable skill interoperability across models.
MCP Memory Deletion:
  • When deletion intent is detected, synchronously trigger deleteMemory and addFeedback.
Improvements
Scheduling & Retrieval Pipeline Refactor:
  • Pipeline refactor: Rebuilt the retrieval chain into four stages—Search -> Enhance -> Rerank -> Filter—scheduled by a unified API.
  • Scheduling optimization: Modularized the task processor and fixed Redis Streams serialization and user_context passing issues.
Skills Improvements:
  • Skill docs localization: Supports local storage of skill docs and generates dedicated URLs, downloadable directly via API.
  • Skill generation optimization: Can now generate richer and more complete skills from users' historical information.
Bug Fixes
Playground Improvements:
  • Fixed known UX issues and improved environment stability.
v2.0.4
New Features
Core Capabilities:
  • Added Skill Memory.
  • Added modules supporting memory versioning.
  • Added chat/completions interface, compatible with OAI protocol.
Improvements
Retrieval Optimization:
  • Reduced duplication between factual and preference memories during retrieval.
Bug Fixes
Bug Fixes:
  • Fixed bug for new users in Playground.
  • Fixed occasional latency when adding memories in cloud service.
  • Resolved graph database efficiency issues.
v2.0.3
New Features
Documentation:
  • Open-source project documentation: Fixed formatting issues, improved example descriptions, resolved Chinese-English inconsistencies, and added practical usage tips.
Dashboard:
  • Knowledge base now supports more file types: JSON, MD, XML.
GitHub Repository:
  • Adjusted the startup approach (introduced a new server router class) and removed the reference to src/memos/product_api.py.
Improvements
Open Source:
  • Optimized text and image memory addition by processing both types of information together as a single message.
  • Knowledge base search now supports the 'all' field.
Bug Fixes
Open Source:
  • Fixed an exception that occurred when the search input text was too long.
v2.0.2
New Features
Released "Knowledge Base Q&A Assistant" Construction Tutorial:
  • Official documentation updated with best practices for "Developing Q&A Assistants Based on Knowledge Bases", guiding you step-by-step to build a "MemOS Knowledge Base Assistant".
Improvements
Knowledge Base Document Retrieval:
  • Further improved the retrieval memory interface's ability to recall detailed knowledge base document content, optimizing answer quality for document-based queries.
Tool Memory Mechanism:
  • Added procedural experience for call trajectories to enhance the auxiliary effect of tool memory;
  • Compressed tool schema information to avoid duplicate addition of ToolSchemaMemory.
Fact Memory Merging and Archiving Mechanism:
  • Added merging and archiving mechanisms in fact memory processing to optimize issues of duplicate memory writing and recall, while ensuring extraction completeness.
Get Memory Interface:
  • Support returning tool memories;
  • Support returning memories under specific filtering conditions for debugging and display scenarios.
Bug Fixes
Get Memory Interface:
  • Fixed call error in get memory interface when parameter include_preference=False.
v2.0.1
New Features
Get Memory Interface Launch:
  • Get memory interface launched, supporting full retrieval of user memories.
Chat Function Launch on Cloud Service:
  • Chat function launched on cloud service, enabling one-click start of "memory-persistent" conversations.
Playground Update:
  • Playground memory management page supports memory deletion, allowing users to manually manage and delete expired memories they no longer need.
Improvements
Delete Memory Interface:
  • Optimized delete memory interface to support deletion of all types of memories, including user memories, knowledge base memories, etc.
Preference Memory:
  • feedback interface added feedback processing operation for preference memory
Memory Retrieval Optimization:
  • search interface supports optional deduplication parameter, supporting semantic similarity deduplication to improve retrieval result diversity.
Bug Fixes
kv cache Issue Fix:
  • Solved compatibility error of kv cache code in new version of memos
Scheduling Module Issue Fix:
  • Fixed error when scheduling enables local mode due to missing redis configuration
v2.0.0
New Features
Knowledge Base & Memory:
  • Added knowledge base support for long-term memory from documents and URLs
Feedback & Memory Management:
  • Added natural language feedback and correction for memories
  • Added memory deletion API by memory ID
  • Added MCP support for memory deletion and feedback
Conversation & Retrieval:
  • Added chat API with memory-aware retrieval
  • Added memory filtering with custom tags (Cloud & Open Source)
Multimodal & Tool Memory:
  • Added tool memory for tool usage history
  • Added image memory support for conversations and documents
Improvements
Data & Infrastructure:
  • Upgraded database for better stability and performance
Scheduler:
  • Rebuilt scheduler with Redis Streams and queue isolation
  • Added task priority, auto-recovery, and quota-based scheduling
Deployment & Engineering:
  • Added lightweight deployment with quick and full modes
Bug Fixes
Memory Scheduling & Updates:
  • Fixed legacy scheduling API to ensure correct memory isolation
  • Fixed memory update logging to show new memories correctly
v1.1.3
New Features
Memory Add & Search:
  • Added async mode (plaintext & preference)
  • Preference Memory now supported
  • Reranker strategy suite
  • BM25 for TreeTextMemory
Scheduler:
  • Modularized API scheduler
  • Redis ORM optimized for history sync & hybrid search
Data & Infra:
  • PolarDB graph backend connection pool/timeout & fixes
  • Unified graph factory (Neo4j/PolarDB/Nebula)
  • Milvus interface & item optimizations
  • Enhanced logging pipeline
  • Nacos-based dynamic configuration
Evaluation:
  • PrefEval field standardization
  • LoCoMo/LongMemEval/PrefEval/PersonaMem evaluation upgrades
Improvements
Plaintext Memory:
  • Standardized preference fields
Framework:
  • Updated API routes
  • More robust error handling & context tracing
Bug Fixes
Scheduler:
  • Fixed query scheduling edge case
  • Corrected message schema inconsistencies
Plaintext Memory:
  • Graph/DB：corrected PolarDB issues
Framework:
  • Fixed SQLite user listing bug
v1.1.1
New Features
Cloud Service 1.0 Beta Release:
  • 🛠️ Ready-to-Use：Deploy directly via cloud API without complex setup
  • 🔄 Cross-Session Memory：Automatically retrieve user profiles, preferences, and behaviors for continuous personalization
  • 📖 Developer-Friendly：Comprehensive documentation and example code support
  • 🎁 Free Quota：Every developer can receive basic quota to explore
Enterprise Group Q&A bot:
  • Supports Q&A based on cloud service knowledge base
v1.0.1
New Features
Group Q&A Bot:
  • Launched group Q&A bot based on MemOS Cube
KV-Cache Performance Optimization:
  • Updated comparative experimental data for KV-Cache on different GPU deployment schemes
  • Optimized test benchmarks and statistical methods
Plaintext Memory Enhancement:
  • Added Reranker sorting functionality for plaintext memory
Playground Updates:
  • Updated Playground version with all the above new features
Improvements
Plaintext Memory Optimization:
  • Optimized plaintext memory hallucination issues
v1.0.0
New Features
Playground:
  • Expanded Playground features and algorithm performance.
MemCube Construction:
  • Added a text game demo based on the MemCube novel.
Extended Evaluation Set:
  • Added LongMemEval evaluation results and scripts.
Improvements
Plaintext Memory:
  • Integrated internet search with Bocha.
  • Added support for Nebula database.
  • Added contextual understanding for the tree-structured plaintext memory search interface.
Bug Fixes
KV Cache Concatenation:
  • Fixed the concat_cache method.
Plaintext Memory:
  • Fixed Nebula search-related issues.
v0.2.2
New Features
Explicit Memory:
  • Implemented internet search integration with Nebula database support
  • Enhanced contextual understanding for memory extraction
  • Deployed internet search interface integrated with memreader for web message processing
KV Cache:
  • Conducted KV Cache evaluation including LMCache comprehensive research and performance testing
  • Completed stress testing for MemOS vllm (v0.2.1) in specific environments and models
  • Refined ttft test data for preloaded KV Cache on Qwen2.5-72B-Instruct model and updated report
  • Completed initial research and design for high-ROI inference service system using domestic hardware
Memory Operation Model:
  • Completed model training evaluation and release
  • Released 4b, 1.7b and 0.6b models supporting memory extraction and integration
Improvements
Documentation:
  • Added English version for first three chapters of Cookbook
Memory Scheduling:
  • Performed code refactoring and functionality improvements
  • Refactored functional modules including monitor, dispatcher and retriever
  • { "Added major code categories": "schemas and utils" }
  • Enhanced network logging functionality
  • Improved scheduler robustness with new exception-catching decorators
  • Implemented locking for shared resources
Development Environment:
  • Updated Docker configuration
  • Updated dim environment configuration
API Enhancements:
  • Added playground context support
  • Enhanced product API functionality
  • Rewrote query module
  • Implemented chat history feature
Bug Fixes
Data Parsing:
  • Fixed date parsing error
  • Fixed memos_w_scheduler example code issues
  • Fixed fine-grained search bug in Nebula graph database
  • Fixed metadata filter retrieval logic
System Integration:
  • Aligned MOSProduct._build_system_prompt signature with MOSCore
Explicit Memory:
  • Fixed general text memory processing logic
  • Fixed memreader component issues
v0.2.1
New Features
MemCube Features:
  • Added plaintext memory + KV Cache functionality with reasoning and decoding performance report
  • Completed development of interactive cube feature with full workflow support and embedding model switching
MemOS System:
  • Released MemOS Neo lightweight version with simplified architecture and core API modules
  • Added MCP support research and preparation work, expanding MCP capabilities
  • Created embedded Agent flow Pipeline documentation with Coze framework integration
Deployment & Environment:
  • Added Docker deployment support
  • Enhanced API layer with multi-model support, compatible with OpenAI/Qwen/DeepSeek and other mainstream models
  • Improved embedding model support
  • Added neo4j Community Edition/Nebular database support
  • Implemented multi-tenant architecture support within single database
Evaluation & Testing:
  • Adapted and evaluated memos API interface format, updated evaluation code
Documentation & Examples:
  • Added new Cookbook content
  • Included Mud game example
v0.2.0
New Features
Website Usability Updates:
  • Added documentation search functionality
  • Implemented Chinese/English documentation switching
  • Enhanced footer navigation links and page editing features
Memory Operator Model:
  • Introduced MemReader-4B small model for memory extraction operations
  • Enabled fully local deployment for restricted network environments
  • Achieved lower cost and faster memory operations with performance exceeding GPT-4o-mini
  • Fine-tuned based on Qwen3-4b with human and model annotation data using supervised fine-tuning
Cross-Platform Framework Adaptation:
  • Added Windows platform deployment support
  • Added Mac platform deployment support
  • Completed adaptation for Linux, Windows, and macOS mainstream operating systems
  • Passed testing on Ubuntu 20.04+/CentOS, Windows 10+/11, macOS 14 Ventura+ platforms
  • Supports stable operation of core modules (memory lifecycle management, MIP protocol, memory cache scheduling)
Playground Progress:
  • Completed engineering, front-end, and algorithm end-to-end connection, and bug fixing and optimization in progress