Developer tests strands-agents and identifies flaws in cross-restart memory and conversational recall

On November 7, 2026, a developer published a hands-on test of strands-agents on DEV Community. The test ran on strands-agents 1.63.0 with Python 3.11, using the amazon.nova-pro-v1:0 model on Amazon Bedrock in us-east-1. Setup steps are documented in the README.md, every number quoted comes from a single real run, and the findings were verified on November 7, 2026.

The central result is a memory paradox. The agent could retain a user's name across a restart, yet still drop that same name seven turns later within the ongoing conversation. Because the failure appears in two separate places, the developer concludes these are two distinct session state management defects rather than one root cause. Fixing one does nothing for the other.

A previous article in the series had established that an agent's state is essentially one list of messages, which frames how conversation content is held and how it can drift. The practical takeaway for anyone building on strands-agents is that durability across restarts does not imply continuity within a conversation, and both problems need to be addressed independently. At the time of writing, the issue remains open: no patch, configuration workaround, or response from the maintainers has been reported.

The developer notes that the reproduction path is fully documented, so others can verify the behavior on the same stack. Until the underlying state handling is revised, anyone relying on long-running conversations should treat in-conversation recall as unreliable even when cross-restart memory appears to work.