Summary

AgentCoreMemorySessionManager.retrieve_customer_context filters retrieved memories using
m.get("relevanceScore", ...), but the MemoryRecordSummary shape returned by
retrieve_memory_records defines the field as score (confirmed via the AWS service model).

The field name mismatch means the fallback value is always used — set to
retrieval_config.relevance_score itself — so the condition relevance_score >= relevance_score
is always True. Every memory passes the filter regardless of the configured threshold.

Root cause

# Before — "relevanceScore" is never present; fallback makes condition always True
memories = [
    m for m in memories
    if m.get("relevanceScore", retrieval_config.relevance_score) >=
    retrieval_config.relevance_score
]

The MemoryRecordSummary boto3 service model shape:

score: Double   ← actual field name

Fix

# After — filter on the correct field; 0.0 default excludes records without a score
memories = [m for m in memories if m.get("score", 0.0) >= retrieval_config.relevance_score]

Using 0.0 as the default ensures records without a score are excluded rather than silently
passed through.

Changes

• session_manager.py: fix field name in relevance filter
• test_agentcore_memory_session_manager.py: update mock memory records to use "score" to
  match the actual API response shape