Changelog

Unreleased

Added

• Google Gemini 1.5 Pro and Flash provider integration
• pgvector native support for PostgreSQL-based vector storage
• Agent-to-agent direct tool call messaging in pipelines
• TypeScript API client (read-only, alpha)

Changed

• BREAKING: memory.episodic.backend config key renamed to memory.episodic.storage — migration: update your config files
• Improved LLM error messages to include model name and request ID for easier debugging

Fixed

• Pipeline checkpoints now correctly serialize large tool outputs over 1MB
• Webhook signature verification now handles UTF-8 request bodies correctly
• Memory pruning no longer accidentally deletes active session entries

v0.9.0 — 2024-11-15

Production Readiness Release

Added

• Docker multi-stage production build (100MB image, non-root user)
• Kubernetes Helm chart with HPA support
• Prometheus metrics endpoint (/metrics) for agent and LLM observability
• Plugin registry: openclaw plugin install/list/remove
• REST API with FastAPI, OpenAPI docs at /docs
• Webhook trigger system with HMAC signature verification
• Response caching with TTL and namespace support
• SQS and Redis Streams queue consumers
• AWS ECS, GCP Cloud Run, and Azure deployment guides

Changed

• Config file format now supports YAML anchors for DRY configuration
• Tool timeout raised from 10s to 30s default (configurable)
• Improved parallel pipeline execution with better resource sharing

v0.8.0 — 2024-09-01

Multi-Agent and Memory Release

Added

• Pipeline orchestration: sequential and parallel multi-agent execution
• Episodic memory with SQLite and PostgreSQL backends
• Semantic memory with Chroma, Qdrant, and Pinecone backends
• RAG indexing pipeline: openclaw rag index
• Anthropic Claude 3 provider integration (Haiku, Sonnet, Opus)
• Ollama provider for local model execution
• Built-in scheduler with cron expression support

v0.7.0 — 2024-07-10

Tool Ecosystem Release

Added

• 20 built-in tools: web_search, read_url, write_file, execute_code, send_email, and more
• Tool sandboxing with configurable permissions
• Playwright browser automation tool
• Python code execution tool with AST-based safety checks
• File operation tools with path traversal protection

v0.6.0 — 2024-05-20

Configuration and Developer Experience

Added

• YAML-first configuration with JSON Schema validation and IDE autocomplete support
• Config init wizard: openclaw config init for guided first-time setup
• Config validation command: openclaw config validate
• Multiple config profile support for dev/staging/production environments
• Python SDK with Agent class for programmatic control
• Structured JSON logging with configurable format and log level

Fixed

• Tool timeout now applies correctly when tools are called inside async contexts
• Config file now reloads cleanly when changed without requiring full restart

v0.5.0 — 2024-04-01

Initial CLI Release

Added

• Core CLI: openclaw run, openclaw start, openclaw stop
• OpenAI GPT-4o and GPT-4o-mini provider support
• 5 built-in tools: web_search, read_url, write_file, read_file, execute_code
• Single-agent task execution with ReAct reasoning loop
• Basic configuration via environment variables
• Token budget enforcement with configurable limits

Reading the Changelog

Each release entry follows the standard Keep a Changelog format with sections for Added (new features), Changed (changes that may require config updates), Deprecated (features that will be removed in a future release), Removed (features removed in this release), Fixed (bug fixes), and Security (security-related patches). When upgrading, focus first on the Changed section — this is where breaking changes are documented.

The full git history is always available at github.com/openclaw-ai/openclaw/commits/main for detailed commit-level change tracking. For significant upgrades spanning multiple minor versions, the MIGRATION.md file in the repository provides step-by-step upgrade instructions.

v0.4.0 — 2024-02-15

Memory and Context

Added

• ChromaDB-based semantic memory (long-term vector search over past interactions)
• Configurable working memory window (limit context to last N turns)
• Memory export and import for agent continuity across sessions
• Anthropic Claude 3 provider integration (Haiku, Sonnet, Opus)

Changed

• Memory configuration moved from top-level to agent.memory namespace
• Default context window limit reduced from no limit to last 20 turns (improves performance)

v0.3.0 — 2024-01-10

Multi-Agent Orchestration

Added

• Pipeline YAML for multi-agent stage definitions
• Parallel pipeline stage execution
• Agent-to-agent context passing between pipeline stages
• Webhook event triggers for pipeline execution

Fixed

• Pipeline stage failures now correctly propagate exit codes
• Memory leaks in long-running pipeline sessions resolved

v0.2.0 — 2023-11-22

Tool Ecosystem

Added

• Plugin system for third-party tool packages
• Built-in tools: send_email, web_search (DuckDuckGo), read_url (BeautifulSoup parser)
• Tool timeout configuration per tool
• Tool usage logging (which tools were called, arguments, return values)

Changed

• Tool definitions migrated from Python decorators to YAML spec + Python implementation

v0.1.0 — 2023-10-05

Initial Alpha Release

• Core ReAct loop (Reason + Act) with OpenAI provider
• Basic file read/write tools
• Single-agent, single-task execution model
• Simple text output (no streaming)

This was the first public alpha. The API was not stable and changed significantly in subsequent releases. We do not recommend running v0.1.x in any environment.

Upgrading Between Versions

The safest upgrade path is always version-by-version rather than skipping multiple minor versions at once. Run pip install openclaw==X.Y.Z for the specific version, test your configuration with openclaw config validate, run a sample task to verify functionality, and only then upgrade to the next version.

If you are running OpenClaw in production with agent configurations that have been tuned to specific behaviors, always test upgraded versions in a staging environment first. LLM-driven behavior can shift when the system prompt formulation or tool-calling logic changes even in patch releases, though we work hard to minimize such changes.

The CHANGELOG.md file in the repository root is always the canonical source of truth. This webpage is generated from that file and may lag briefly behind new releases. For the most current information, refer to the repository directly.

Stay Updated

Subscribe to GitHub releases by clicking the Watch button on the repository and selecting Releases Only. All releases are also announced in the project Discord and on the official social media channels. Security-related releases include immediate notifications to all registered users with active API keys, so you are never caught unaware by a critical patch.