.NET: Python: MemoryFileStore topic slug collisions silently overwrite another topic's memories #9723

Description

MemoryFileStore keys every topic file by _slugify_topic(topic), which strips everything that is not a to z or 0 to 9. The mapping is many to one: C ++ and C # both slug to c, node.js and node js both slug to node-js, and API and api collide too. write_topic then does an unconditional overwrite of the slug path, so writing a second topic that shares a slug with an existing one silently destroys the first topic's memories.

Reproduced on current main: write topic C ++ with the memory Modern C ++ prefers smart pointers, then topic C # with C # 12 primary constructors. Afterwards get_topic("C ++") returns the C # memory, the C ++ memory is gone, and list_topics shows a single topic.

The same file family deliberately uses the injective _storage_key_segment derivation for owner ids, session ids, source ids, and scopes, precisely so two byte distinct values do not share a storage location. The topic filename is the one mapping left below that standard.

Expected Behavior

Distinct topics must never share a memory file. Writing C ++ and C # should persist two files whose contents survive, and get_topic by topic name should keep returning each topic's own memories.

Steps to Reproduce

1. Create a MemoryFileStore and write two topics whose names share a slug.
2. Read the first topic back.

Minimal Reproduction

store.write_topic(session, MemoryTopicRecord(topic="C++", summary="cpp", memories=["smart pointers"], updated_at=...), source_id="memory")
store.write_topic(session, MemoryTopicRecord(topic="C#", summary="csharp", memories=["primary constructors"], updated_at=...), source_id="memory")
store.get_topic(session, source_id="memory", topic="C++").memories  # returns the C# memory on main

Package Versions

agent_framework core from main.

Python Version

3.11