2025-07-15 - Mastra Release
CLI / Playground
- Add working memory feature to the playground, enabling users to view and edit their working memory while interacting with an agent. #5665 [IMPORTANT]
- The ability to display reasoning in the playground interface. #5829
- The agent tables by extracting several icons into reusable components. #5843
- Add a Link component to enable shared navigation logic between core, cloud (Next.js), and playground (React Router) environments despite their differing Link APIs. #5844
- CLI templates #5927 [IMPORTANT]
- Refactored the playground UI to export the network table for use in the cloud environment. #5932
- Correctly handle navigation to '/new' as a threadId, ensuring 'new' is not incorrectly used as the threadId. #5946
- Correctly route agent network requests to the vNext AgentNetwork instead of defaulting to the legacy network. #5950
Client SDK - JS
- Fix usage of crypto.randomUUID in the client SDK to address compatibility or dependency concerns. #5853
RAG
- Remove cohere-ai dependency in favor of a direct fetch call, deprecates the CohereRelevanceScorer class in @mastra/core, and resolves a peer dependency issue with Cloudflare Workers. #5819
- Export types to make them available for external use. #5867
- Fix semantic recall system message formatting in @mastra/core by including dates, times, and thread information, resulting in significantly improved longmemeval scores. #5916
Deployer
- Support for handling private packages as external dependencies in the mastra build bundle output. #5834
- Toggle installDependencies step for the Cloudflare deployer build process. #5903
Memory
- Scripts to prepare and run benchmarks for longmemeval data across various memory configurations, enabling evaluation of memory features but does not include the actual data or results. #5925 [IMPORTANT]
MCP
- Fix bug in @mastra/schema-compat that caused a 'Cannot read properties of undefined (reading 'typeName')' error. #5942
Memory [IMPORTANT]
- Assistant message merge logic, ensures merged messages are consistently updated and persisted, and removes unnecessary step-start normalization for new assistant messages. #5866
- Per-resource memory settings are correctly passed into updateWorkingMemory() in the agent tool, preserving the intended memory scope. #5913
- Support for passing a JSON schema directly to the working memory configuration in the core and memory modules. #5918
- An experimental 'vnext' version of working memory with improved update logic and new arguments to reduce accidental memory erasure during long, multi-threaded interactions, resulting in a 20% performance boost but still requiring further testing before full adoption. #5924
Networks
- Pipes the runtimeContext to vNext network agents, the client SDK, and the playground, ensuring consistent runtime context handling across these components. #5828
- Fix thread creation if it does not already exist in the agent network stream, generate, and loopStream functions, addressing a bug. #5905
Storage
- Ensure the resourceId is selected when retrieving messages. #5824
Voice
- OpenAIRealtimeVoice can now properly handle errors returned from OpenAI. #5653
Workflows
- The workflow foreach loop emits the full result at once instead of emitting an event for each loop iteration. #5791
- Workflow status is only set to success after all steps complete successfully, adds related test cases, and tests error handling for resuming non-suspended workflows or steps. #5847
- Fix resuming a nested workflow within a loop workflow was causing failures. #5808
- RuntimeContext is saved in snapshots and made available during step executions, allowing workflows to retain and access runtimeContext after resuming from suspension. #5911
- That removed runtimeContext values are not saved in snapshots and adds corresponding test coverage. #5933