Summary

Add Mem0 as a new memory store type alongside the existing simple, file, and blackboard stores. Mem0 provides cloud-managed semantic search and graph-based memory, enabling agents to remember user preferences and past interactions across sessions without requiring local embedding computation or persistence.

Motivation

Currently, all memory stores are local — they reset between runs (simple), require local disk and embeddings (file), or are append-only within a session (blackboard). There is no option for persistent, cloud-managed memory that works across sessions and devices.

Mem0 fills this gap by providing:

• Semantic search over stored memories (server-side embeddings)
• Graph relationships between entities for richer retrieval
• Cross-session persistence without local file management
• Sub-50ms retrieval latency

Proposed Changes

• Add Mem0MemoryConfig dataclass following the existing BaseConfig pattern with FIELD_SPECS for UI generation
• Add Mem0Memory(MemoryBase) implementation using the mem0ai Python SDK with lazy imports
• Register mem0 as a new store type via the existing plugin registry in builtin_stores.py
• Add mem0ai as a dependency
• Add unit tests and a YAML example workflow

Key Design Decisions

• Uses mem0ai Python SDK — not raw API calls
• Lazy import — mem0ai is only imported when a Mem0 store is actually instantiated, not at startup
• Correct entity scoping — user_id and agent_id use filters dict for search (with OR filter when both are configured) and top-level kwargs for add
• Agent memories use role: "assistant" — matches Mem0's entity-scoped memory model
• load()/save() are no-ops — Mem0 manages persistence server-side

Configuration Example

memory_stores:                                            
  - name: agent_memory
    type: mem0
    config:
      api_key: ${MEM0_API_KEY}
      agent_id: my-agent