PRD-39: Memory Migration to Mem0

Status: 🟡 Planning Phase Date: January 27, 2026 Author: Automatos AI Related: PRD-05 (Memory Systems)

---

Executive Summary

The objective is to migrate Automatos AI's internal memory system to use mem0, a dedicated memory layer for AI agents. Mem0 offers a robust, open-source solution for managing long-term user memory, session context, and intelligent retrieval. We will utilize the self-hosted mem0 instance running on Railway (automatos-mem0-server.railway.internal) to provide a unified memory backend for all Automatos consumers (Chatbot, Workflows, Agents).

This migration aims to resolve issues with memory flakiness ("calling too often", "saying randomly") by leveraging mem0's production-ready features like intelligent fact extraction and structured retrieval.

Goals

1. Migrate Memory Backend: detailed in modules/memory to use mem0 as the primary storage and retrieval engine.

2. Unified Memory: Ensure all consumers (Chatbot, Agents) read/write to the same mem0 instance.

3. Improved Stability: addressing existing reliability issues by adopting mem0's proven architecture.

4. Simplicity: Reduce custom code maintenance by offloading memory complexity to mem0.

Architecture

Current State

• Storage: pgvector (Vectors) + Custom Knowledge Graph.

• Logic: Custom MemoryInjector, HierarchicalMemorySystem.

• Issues: Retrieval logic is custom and sometimes overly aggressive or irrelevant.

Proposed State

• Backend: Mem0 Server (Railway Enclave).

• Client: mem0ai Python SDK (or direct API if SDK is too heavy/incompatible).

• Integration Point: modules/memory/integrations/mem0_client.py.

• Pattern: Hybrid approach initially (as recommended in research), but moving towards full mem0 adoption.

Flow:

1. Chatbot receives message.

2. Memory Service calls mem0 v1/memories/search (via SDK/API) with user_id (scoped to Workspace/User).

3. Context Injection: Relevant memories injected into LLM context.

4. Storage: Post-response, conversation is sent to mem0 v1/memories/ (add) for async processing/fact extraction.

Implementation Details

1. New Client Integration

Create modules/memory/integrations/client.py:

• Wrapper around mem0 SDK or requests.

• Handles configuration (URL, API Key).

• methods: add(), search(), get_all().

2. Refactor Memory Service

Update modules/memory/service.py:

• Replace/Augment HierarchicalMemorySystem calls with Mem0 client.

• Ensure user_id mapping: f"workspace_{workspace_id}_user_{user_id}" to ensure isolation.

3. Dependencies

• Add mem0ai to requirements.txt.

• Env Var: MEM0_API_URL=http://automatos-mem0-server.railway.internal.

4. Verification

• Use scripts/test_mem0_railway.py for connectivity checks.

• New Integration Test: tests/integration/test_memory_flow.py.

Migration Strategy

1. Add Mem0 Client: Parallel to existing memory.

2. Dual Write: Write to both systems (optional, or just switch).

3. Switch Read: Toggle to read from mem0.

4. Deprecated: Remove old memory code after validation.

Success Metrics

• Connectivity: 100% success rate connecting to Railway instance.

• Latency: Retrieval < 200ms.

• Relevance: Subjective improvement in chat context recall.