fix: insert retrieved LTM before last user message to avoid prefill error

[notgitika]

Summary

• Fixes assistant-prefill error on Claude 4.6+ models (Opus 4.6, Sonnet 4.6) when using AgentCoreMemorySessionManager with LTM retrieval

• Previously, retrieved memory was appended as an assistant message after the last user message, making it the final message sent to the model. Claude 4.6 rejects this with: "This model does not support assistant message prefill. The conversation must end with a user message."

• Now inserts the assistant memory message before the last user message, keeping the user's query last

• For the first message in a conversation (where prepending an assistant message would violate Bedrock's "must start with user" rule), injects memory as a content block within the user message instead

• Renames <user_context> tag to <retrieved_memory> for semantic accuracy

Closes #191

Related issue strands-agents/sdk-python#1694

Testing

• Ran all unit tests
• Reproduced the prefill error against Opus 4.6 with the old code path (ValidationException: does not support assistant message prefill)
• Verified both fix paths (insert-before and inline-in-first-msg) work against Opus 4.6
• Full end-to-end test with AgentCoreMemorySessionManager + Opus 4.6 — no prefill error
• Integration tests updated (<user_context> → <retrieved_memory>)

[jsgsamzn]

If I'm reading this correctly:

1. If there's only one user message, the memory is placed in a content block in that initial user's message.
2. If there's more than one message, the memory is placed in a new "assistant" message before the latest user message.

Is there a potential issue with the role not being consistent here?

[notgitika]

Is there a potential issue with the role not being consistent here?

You're right, there could be. I've jsut updated it to always inline as a content block in the last user message so the role is consistent.

[vivdalal]

LGTM

[notgitika]

Related issue: strands-agents/sdk-python#1736