Agentic Memory

A Mem0-inspired long-term memory system for chat, rebuilt from scratch with DSPy ReAct agents, OpenAI embeddings, and Qdrant.

Agentic Memory studies the Mem0 paper and reimplements its core loop: extract facts from a conversation, store them as vectors, retrieve them later, and decide ADD, UPDATE, DELETE, or NOOP as the conversation changes. A response agent answers from semantic search, and a second agent writes memory only when the user states something worth keeping.

Tech stack

PythonDSPyQdrantOpenAIReAct

Features

Dual DSPy ReAct agents for memory-aware replies and memory writes
Semantic search over user-scoped, category-tagged facts in Qdrant
ADD, UPDATE, DELETE, and NOOP operations chosen by the memory agent
OpenAI embeddings for retrieval and durable fact storage
Separate session transcript and long-term memory that survives restarts

Challenges

Letting the memory agent choose ADD, UPDATE, DELETE, or NOOP instead of hardcoding the branch
Keeping retrieval scoped to a user while still finding the right fact among similar memories
Mapping retrieved list indexes back to Qdrant point IDs for updates and deletes

Learnings

Rebuilt Mem0’s memory architecture from the paper using two coordinated ReAct agents
Separated ephemeral chat history from durable vector memory so recall works across sessions
Used category tags and user filters to keep semantic search precise

Github