Describe the bug

During load testing in #837, profiling shows heavy lock contention in EpisodicMemoryManager and low requests-per-second in locust, even with a low number of users. The problem is worse as we add more users and load.

Steps to reproduce

Use the locustfile.py and profile_memmachine.sh from #837 to generate the load and profile the FastAPI processes running inside the memmachine-app Docker container.

Expected behavior

• Remove serialization locks in high contention code paths.
• Expect higher Requests per Second per FastAPI worker with Locust.

Environment

• OS: Ubuntu Linux 25.04
• MemMachine: v0.2.2
• Locust load generator

Additional context

The open_or_create_episodic_memory method currently holds a global lock while performing database I/O to fetch session info. This serializes all requests to that worker.

Suggested Solution: Use a double-checked locking pattern or a per-session lock to allow concurrent I/O for different sessions.