Big news this week! We've achieved state-of-the-art performance on the LongMemEval benchmark with our memory improvements, reaching 80% accuracy. Plus, we're shipping new CLI templates, reasoning display in the playground, and major improvements across the board.
Let's dive in:
Memory: State-of-the-art LongMemEval Performance
We implemented the LongMemEval benchmark and achieved 80% accuracy — outperforming other specialized memory solutions. This breakthrough came entirely from improvements to semantic recall.
Improved Semantic Recall for SOTA Performance
We restructured how recalled messages are presented to LLMs, achieving state-of-the-art results:
- Messages are now grouped by date with clear time labels
- Better distinction between current and past conversations
- Improved formatting that includes contextual thread information
These changes alone brought us to 80% accuracy on LongMemEval — proving that RAG absolutely works for agent memory when implemented correctly. Read our deep dive on using RAG for agent memory.
vNext Working Memory Updates
As a bonus improvement from our research, we've shipped an experimental 'vNext' version of working memory with smarter update logic:
1// Enable the improved memory tool
2const memory = new Memory({
3 provider: myProvider,
4 options: {
5 workingMemory: {
6 enabled: true,
7 template: `Track user preferences and key details`,
8 version: "vnext", // Available now!
9 },
10 },
11});The new API prevents accidental memory erasure during long conversations. Instead of wholesale replacements, agents now provide update reasons and can make targeted changes — a nice quality-of-life improvement for working memory users.
CLI & Playground: Major Usability Upgrades
Reasoning Display
The playground now shows agent reasoning in real-time. Watch your agents think through problems step-by-step as they work — invaluable for debugging and understanding agent behavior.
Working Memory Improvements
You can now view and edit working memory directly in the playground while interacting with agents. No more guessing what your agent remembers — see it and modify it live.
We also added support for passing JSON schemas to working memory.
CLI Templates
We've improved the CLI with better template support in the create-mastra command. These templates provide pre-configured starting points for common use cases, complete with example agents, tools, and workflows tailored to specific scenarios. Create templates with:
npx mastra@latest create --templateWhen you run create-mastra, you can now choose from different setups that include the right dependencies and boilerplate code for your project type. This makes it even faster to get started with Mastra.
Coming soon: We're working on a library of ready-to-use Mastra templates for specific use cases. Stay tuned for the next release!
Other Important Updates
Networks
- Agent networks now properly handle runtime context across all components
- Fixed thread creation bugs in network
stream,generate, andloopStreamfunctions - Networks correctly route to vNext AgentNetwork instead of legacy versions
Workflows
- Added runtime context persistence in workflow snapshots — workflows retain context after resuming from suspension
- Fixed nested workflow execution within loops
- Workflow status correctly reflects completion state
Deployer
- Support for private packages as external dependencies
- Toggle for installDependencies in Cloudflare deployer
Voice
- OpenAI Realtime Voice now properly handles error responses from OpenAI
RAG
- Removed cohere-ai dependency in favor of direct fetch calls
- Fixed semantic recall system message formatting to include dates, times, and thread information
That's all for this week! We're especially excited about the memory improvements — 80% on LongMemEval proves that RAG absolutely works for agent memory when done right.
For the complete list of changes and fixes, check out the full release notes.
Happy building! 🚀