Skip to content
Agents tracked: 501 Releases (7d): 777 Agent status: no reported issuesSince last limit reset: Claude 10d · Codex 8d For AI agents: Markdown Updated Nov 21, 2026

MemOS changelog: what's new each month

Every stable MemOS release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 12 months covered; the current month updates daily.

November 2026 · so far

1 release: 2.0.36

November brought stability improvements and rate limiting controls for MemOS. The release focused on fixing memory handling and expanding configuration options for LLM operations.

Highlights

  • Per-model Redis GCRA rate limiting is now configurable for fine-grained control over LLM request rates.
  • Memory recall in the playground now searches across all cubes instead of a limited subset.
  • OpenClaw installation and automatic memory hooks have been restored after a regression.

New

  • Configurable per-model Redis GCRA rate limiting for LLM providers

Improved

  • Memory recall in the playground now spans every searched cube
  • Memory evolution and Hermes installation are now more robust
  • Plugin stability and reliability enhancements

Fixed

  • LLM facade no longer drops options before passing to providers, restoring per-operation provider switches
  • OpenClaw installation process restored
  • Automatic memory hooks now function correctly
  • Plugin memory evolution and Hermes installation issues resolved

October 2026

1 release: 2.0.21

October brought deeper configuration control for local models and improved memory stability. Runtime requirements were updated to support the latest Node versions.

Action needed
  • Node 22 or newer required (previously supported older versions)

Highlights

  • Local model settings are now configurable through the Viewer interface, including mirror URLs, cache directories, and offline mode.
  • Memory recovery and capture handling are now more stable, with fixes for orphan traces and LLM response retries.
  • Node 22 or newer is now required for MemOS 2.0.21 and later.

New

  • Mirror URL configuration for local models in Viewer settings
  • Model path exposure in Viewer settings page
  • Cache directory configuration option
  • Offline mode toggle for local embeddings
  • Load timeout control for local embeddings

Improved

  • Dirty-data recovery process for memory integrity
  • Orphan trace write handling
  • LLM empty-response retry stability

Fixed

  • Dirty-data recovery failures
  • Orphan trace writes
  • LLM empty-response retry instability

September 2026

5 releases: 2.0.18 → 2.0.34

September brought stability improvements across MemOS's core memory systems and installation processes. The month focused on fixing recovery mechanisms and reducing connection friction.

Highlights

  • Fixed dirty-data recovery and orphan trace writes to improve memory reliability
  • Enhanced Hermes and Bridge experience to reduce startup and connection friction
  • Restored OpenClaw installation and automatic memory hooks
  • Improved memory evolution and Hermes installation reliability
  • Fixed preference memory and SkillMemory extraction

Improved

  • Memory evolution reliability and performance
  • Hermes installation process
  • Bridge-flow handling
  • Local plugin release notes resilience for large releases
  • Numeric local-plugin source reference normalization

Fixed

  • Dirty-data recovery in memory systems
  • Orphan trace writes
  • LLM empty-response retry stability
  • OpenClaw installation process
  • Preference memory handling
  • SkillMemory extraction links

August 2026

7 releases: 2.0.13 → 2.0.32

August brought stability improvements across MemOS with enhanced Windows runtime support, better Hermes integration, and new DeepSeek Harness integration for long-term memory access. The month focused on fixing resource management issues and improving plugin reliability.

Highlights

  • DeepSeek Harness integration now enables automatic memory recall for MemOS Local installations with configurable agent profiles.
  • Fixed foreground resource starvation issues and added Retry-After header support for better resource handling.
  • Improved Windows runtime lifecycle management for Hermes view daemon startup and status probing.
  • Enhanced installer now automatically handles missing pnpm dependency during one-click DSH installation.
  • Resolved embedding stability issues and improved handling of large embedding inputs to prevent crashes.

New

  • DeepSeek Harness memory adapter for automatic recall integration.
  • One-click installation support for MemOS Local with DeepSeek Harness profiles via install.sh.
  • Automatic pnpm dependency preparation during installation when not found in PATH.
  • Enhanced native dependency installation for DSH Linux with optimized CUDA handling.

Improved

  • Recall relevance and host input handling.
  • Provider startup wait times and link management in Hermes integration.
  • Embedded input processing to handle very large inputs without crashes.
  • Fault tolerance for missing developer peer conflicts.
  • Lock detection and PID management to avoid locking conflicts.
  • Stale lock detection accuracy through improved pidIsAlive checks.
  • Consistency between source code layout and npm package distribution for DSH Viewer assets.

Fixed

  • Foreground resource starvation preventing proper resource allocation.
  • Windows Hermes view daemon startup and status probe failures.
  • SimpleTextSplitter fallback URL-safety issues.
  • Embedding review validation problems affecting plugin stability.
  • DSH Viewer asset resolution in clean installations and release packages.
  • Cached embedder factory instances not preserving configuration.

July 2026

8 releases: 2.0.22 → 2.0.27

July brought substantial improvements to MemOS memory processing, retrieval, and plugin stability. The month focused on optimizing performance for large datasets, enhancing Chinese language support, and fixing critical bugs across multiple components.

Highlights

  • Vector scanning performance optimized for large-scale data with restructured batch processing logic.
  • Chinese keyword recall enhanced with improved CJK word segmentation for better retrieval.
  • Lightweight memory mode configuration now controllable to reduce background processing overhead.
  • L3 abstract model now has dedicated LLM configuration for independent management of summarization.
  • Hermes integration stability improved with better provider startup and connection management.

New

  • Lightweight memory mode configuration to control memory evolution flow.
  • Dedicated L3 LLM configuration entry for abstract summarization stage management.
  • Chinese keyword recall support with enhanced CJK word segmentation.
  • Batch reflection scoring functionality for performance optimization.
  • Yunxiao sync preflight check.

Improved

  • Vector scanning performance optimized through restructured batch processing logic.
  • Retrieval context boundaries clarified with XML organization to reduce model misinterpretation.
  • Hermes integration stability enhanced with improved provider startup and installation scripts.
  • API documentation clarified for product model descriptions.
  • CI workflow updated for better stale branch management.

Fixed

  • Preference memory prompt over-extraction issue.
  • Windows path format mismatch causing better-sqlite3 loading failure.
  • FTS5 query failures with Hermes Agent instructional prompts.
  • Local viewer environment variable parsing in configuration files.
  • OpenAI-compatible endpoint detection instability.
  • ShareCubeWithUser function passing swapped arguments to access validation.

June 2026

2 releases: 2.0.19 → 2.0.20

June brought bug fixes and documentation improvements across MemOS. The month focused on stability and accessibility through translations.

Highlights

  • Fixed a critical issue where autoInstall setting was being bypassed by the installRecommended path.
  • Corrected task failure recording that was incorrectly marking failed tasks as successful experiences.
  • Expanded documentation with Chinese and English translations for core API and evaluation resources.
  • Improved search pipeline with new context rendering hooks for better flexibility.

New

  • Added harness feature to MemOS.

Improved

  • Search pipeline now includes hooks for customizing context rendering.
  • OpenClaw runtime startup now has proper guards to prevent errors in the memos-local-plugin.

Fixed

  • autoInstall configuration setting was not being respected; installRecommended was bypassing the setting.
  • Failed tasks were incorrectly recorded as successful experiences in the system.
  • OpenClaw runtime startup issues in the memos-local-plugin.

May 2026

3 releases: 2.0.15 → 2.0.17

May brought a major architectural refresh to the memos-local-plugin with the new Reflect2Evolve framework, alongside improvements to Windows support and various bug fixes across the platform.

Action needed
  • Memos-local-plugin web/ directory renamed to viewer/; unused site/ scaffolding removed

Highlights

  • Memos-local-plugin completely rewritten with Reflect2Evolve architecture for mid-turn reasoning capture
  • New one-round-one-card UI with language-aware knowledge and reasoning boundary prompts
  • Enhanced Windows installation experience for memos-local-plugin
  • Stage log and upload skill logic added to core functionality
  • Embedding dimension inference now works correctly during rebuilds

New

  • Stage log capability
  • Upload skill logic
  • Memos-local-plugin v2 with Reflect2Evolve architecture
  • Mid-turn reasoning capture in memos-local-plugin
  • Language-aware knowledge in memos-local-plugin
  • L2/L3 boundary prompts for better reasoning scaffolding

Improved

  • Dispatcher task optimization
  • Windows installation experience for memos-local-plugin
  • Environment configuration with .env.example-full template
  • ARMS telemetry coverage in memos-local-plugin v2
  • Embedding dimension inference for plugin rebuilds

Fixed

  • Chunking and PolarDB metadata bug
  • Embedding dimension inference during memos-local-plugin rebuilds
  • ARMS telemetry coverage gaps in memos-local-plugin v2

April 2026

3 releases: 2.0.12 → 2.0.14

April brought major architectural improvements to MemOS, including a complete rewrite of the local plugin system and enhanced memory sharing capabilities. The month focused on reliability, plugin performance, and better context handling.

Highlights

  • Complete rewrite of the local plugin system with new Reflect2Evolve architecture for improved reasoning capture
  • Enhanced memory sharing with better scoping and owner management across hub instances
  • Improved hub embedding and parallel recall with auto-update reliability fixes
  • New stage logging and better environment configuration examples
  • Optimized dispatcher task performance

New

  • Stage logging capability for better visibility into process execution
  • Context engine slot registration for local plugins
  • .env.example-full configuration template for reference
  • Upload skill logic for extending plugin capabilities
  • Mid-turn reasoning capture in Reflect2Evolve plugin architecture

Improved

  • Memory sharing now scoped by hubInstanceId with improved owner handling
  • Hub embedding and parallel recall performance enhanced
  • Dispatcher task optimization for better efficiency
  • Auto-update reliability improved for local OpenClaw plugin
  • Cloud plugin code updated and improved
  • Package installation now uses npm instead of npx

Fixed

  • Chunking and polardb metadata bug resolved
  • SecretRef apiKey resolution in OpenClaw fallback
  • Sharing state scoping and owner assignment in memos-local
  • Environment configuration example corrections

March 2026

4 releases: 2.0.8 → 2.0.11

March brought performance optimizations, expanded LLM configuration options, and improvements to memory handling across MemOS. The month saw four releases focused on stability and extending the platform's capabilities.

Highlights

  • Knowledge base chunk logic optimized for better performance
  • New general LLM configurations for MemReader with model training support
  • Multi-OpenClaw sharing and dual-instance isolation added for local deployments
  • Memory storage and preference handling improved to prevent data duplication
  • Scheduler and MemCube now use general LLM by default for consistency

New

  • Auto-cleanup worker for working memory management
  • Image and file info tracking capabilities
  • General configs and training model support for MemReader
  • Backup configuration option for OpenAI MemReader
  • OpenClaw local plugin version (memos-local-openclaw)
  • Multi-OpenClaw sharing with dual-instance isolation
  • Install scripts for easier deployment
  • Deepseek provider support

Improved

  • Knowledge base chunk logic optimized
  • PolarDB ThreadedConnectionPool optimized
  • Export graph functionality improved
  • Image LLM in file-content set as default model
  • Preference memory now uses relativity instead of score
  • Thread pool optimized
  • Auto-recall migrated from beforeagentstart to beforepromptbuild
  • MemReader callback image LLM parser redirected to general LLM

Fixed

  • Activation memory config crash with OpenAI when calling getdefault()
  • Unicode sanitization for cloud embedders
  • File names being stored as part of memory
  • Scheduled tasks not using Redis
  • Windows directory path error
  • Zhipu thinking mode causing summarizer output truncation

February 2026

3 releases: 2.0.5 → 2.0.7

February brought significant database and search enhancements to MemOS, including PostgreSQL vector support and improved full-text search capabilities across multiple database backends. The month focused on stabilizing memory management features and expanding query flexibility.

Highlights

  • Added PostgreSQL with pgvector as a new graph database backend option.
  • Implemented keyword search and full-text search capabilities for Neo4j databases.
  • Enhanced memory filtering to support time-based conditions including hours, minutes, and seconds.
  • Improved memory tagging system to include keywords related to content.
  • Added returnfields parameter to search methods for customizable result formatting.

New

  • PostgreSQL + pgvector graph database backend support.
  • Keyword search functionality for memory queries.
  • Full-text search for Neo4j database backend.
  • Delete node by memory cube ID and recover memory by memory cube ID operations.
  • Time-based filtering with hours, minutes, and seconds support.
  • Returnfields parameter for customizing search result fields.
  • Toggle for full-text retrieval path selection.

Improved

  • Memory version management with new data structures and classes.
  • Parallel processing for large memory items with chunk splitting.
  • Database compatibility for username retrieval across all database types.
  • Memory filtering to support both field conditions and logical operators together.
  • LLM call functionality with thinking feature support.
  • Username and keyword optimization in logging.
  • Tag system now captures content-related keywords more effectively.

Fixed

  • File URL no longer incorrectly added to memory values.
  • PolarDB parameter configuration bug.
  • Graph edge addition error.
  • LLM call enable thinking bug.
  • Score field handling for preference memory (replaced relativity with score).
  • Log judgment scoping by username in add/update operations.

January 2026

4 releases: 2.0.1 → 2.0.4

January brought refinements across MemOS's core systems, with focus on stability, knowledge management, and chat functionality. The month included scheduler improvements, enhanced Neo4j support, and several fixes to playground and API behavior.

Highlights

  • Added dehallucination operation to the scheduler for more reliable agent reasoning.
  • Improved Neo4j Community Edition support and MCP environment loading.
  • Refactored examples with unified structure and updated demonstration materials.
  • Enhanced knowledge base handling with raw text adoption and improved retrieval backtracking.
  • Added skill memory capability for agents.

New

  • Export graph data page for visualizing knowledge structures.
  • Recaller feature for faster add process.
  • Skill memory system for agents.
  • Model error traceback for debugging.
  • Openwork integration.

Improved

  • Optimized cloud service API.
  • Enhanced chat playground stability and handler initialization.
  • Improved LLM robustness through unified image analysis prompts and context injection.
  • Refined rerank configuration to handle longer inputs.
  • Better scheduler performance and fine-tuning APIs.
  • Enhanced feedback mechanism in preferences.

Fixed

  • Chat playground bugs and initialization issues.
  • Knowledge retrieval backtracking logic.
  • JSON loading errors.
  • Playground chat functionality.
  • Scheduler component initialization to prevent repeated scheduler creation.
  • API examples and documentation.

December 2025

1 release: 2.0.0

December brought MemOS 2.0.0 with significant enhancements to memory management and worker capabilities. The release introduces new scheduling features and addresses key operational issues.

Highlights

  • New top-k filtering for working memory to improve performance
  • Redis queue integration for enhanced scheduler functionality
  • Simplified tree structure for better efficiency
  • Deduplication of memory to reduce redundancy
  • Max worker configuration for resource control

New

  • Top-k selection for working memory
  • Redis queue support in scheduler
  • New API search functions for mixture mode
  • New API search functions for fine mode
  • Max worker setting

Improved

  • Simplified tree implementation
  • Memory deduplication
  • Scheduler API capabilities

Fixed

  • Chat timestamp correction (was showing 2023 instead of 2025)

Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. MemOS: pricing, features and alternatives · All changelogs