Changelog

New Features:

• Model-Level Retries: Improved retry logic added to the model execution layer. That means setting retries=n on your model, would retry directly at the model execution layer, which is useful for dealing with ratelimits imposed by the model provider. Agent-level retries still exist and now apply to any exceptions raised in the agent execution loop. See the docs.

Improvements:

• Propagate context variables to parallel workflow step threads: Fixed context variable propagation to child threads in parallel workflow execution. Without this fix, contextvars (used by tracing/instrumentation libraries like OpenTelemetry) are not propagated when steps run in parallel using ThreadPoolExecutor.

Bug Fixes:

• Knowledge Retriever Update: Fix incorrectly passing dependencies instead of run_context on aget_relevant_docs_from_knowledge

Removals

• MemoriTools has been removed by the Memori team, since it is now obsolete given how the Memori framework will work going forward. See the new docs.

What's Changed

• [feat] update memorisdk to v3.0.5 by @Boburmirzo in #5607
• fix: Pass dependencies to custom knowledge retrievers by @willemcdejongh in #5621
• chore: tracing debug logs cleanup by @kausmeows in #5624
• feat: Add model level retries by @dirkbrnd in #5596
• fix: propagate context variables to parallel workflow step threads by @RogerHYang in #5579
• chore: Release 2.3.8 by @dirkbrnd in #5625

New Contributors

• @RogerHYang made their first contribution in #5579

Full Changelog: v2.3.7...v2.3.8