perf: InMemoryFileStore LRU cache is unbounded per conversation

[csmith49]

Problem

Each LocalConversation gets its own InMemoryFileStore with an LRU cache. The cache has no configured maximum size. With N conversations on a server, each accumulating event data, the total memory from file store caches grows without bound.

Location: io/memory.py

Evidence

Code review. Each event is stored as a JSON string in the file store. A conversation with 2000 events at ~500 bytes each = ~1 MB per conversation in the cache alone. With 500 conversations: ~500 MB just in caches.

Severity: High

Impact of fix

Bounded memory per conversation, predictable server memory usage.

Suggested fixes

• Configure a max cache size per store
• Or use a shared LRU across conversations with a global memory budget

Related issues and PRs

• [FileStore]: fix Use caching on demand to reduce memory usage #1850 — [FileStore]: fix Use caching on demand to reduce memory usage (closed PR, directly addressed this)
• feat(sdk/io): add PostgreSQLFileStore for persistent EventLog #2902 — feat(sdk/io): add PostgreSQLFileStore for persistent EventLog (open PR)
• Add PostgreSQLFileStore — persistent EventLog for ephemeral deployments #2903 — Add PostgreSQLFileStore — persistent EventLog for ephemeral deployments (open issue)
• feat(sdk): allow injecting a custom FileStore into ConversationState #2509 — feat(sdk): allow injecting a custom FileStore into ConversationState (open PR)

Discovered during profiling investigation (code review), May 2026. openhands-sdk v1.19.1.