Agentic Memory Loop

An exercise in implementing agentic memory — built as prep for MLADS+ Memory (Feb 2026).

The agent uses Azure OpenAI to maintain long-term memory across sessions via embedding-based recall, semantic fact extraction, and memory consolidation.

How it works

1. Recall — embeds user input and searches past memories by cosine similarity
2. Act — generates a response using relevant memories + conversation history
3. Store — saves the interaction as episodic memories with embeddings
4. Extract — distills durable facts (semantic memory) from conversations on exit
5. Forget — decays old memories and consolidates redundant ones

Setup

pip install openai python-dotenv numpy

Create a .env file:

AZURE_OPENAI_ENDPOINT=your_endpoint
AZURE_OPENAI_KEY=your_key

Run

python agent.py