Engineering Short & Long Term Agent Memory using MongoDB

A hands-on lab exploring how to build agents that remember — within a session, across sessions, and at scale — using MongoDB as the memory backend.

What you'll build

Notebook	Memory type	Key primitive
01-short-term-memory.ipynb	Session / Working memory	MongoDBSaver (LangGraph checkpointer)
02-long-term-memory.ipynb	Episodic + Semantic memory	VoyageAI embeddings + $vectorSearch
03-semantic-cache.ipynb	Semantic Cache	VoyageAI embeddings + $vectorSearch

Concepts covered

Short-Term Memory
├── Working memory    — LLM context window
└── Session memory    — LangGraph checkpoints in MongoDB (MongoDBSaver)

Long-Term Memory
├── Episodic          — records of past interactions
├── Semantic          — facts, preferences, entity memory
└── Procedural        — (referenced in slides)

Semantic Cache       — skip the LLM for semantically equivalent queries

The full memory pipeline: Aggregate → Encode → Store → Organise → Retrieve

Prerequisites

• GitHub Codespaces (recommended) or Docker + VS Code Dev Containers
• A VoyageAI API key — used for embeddings
• An Anthropic API key — used for the LLM

Getting started

1. Open the repo in a Codespace (or Dev Container).
2. Add your secrets in Codespace settings (or a .env file for local):
   • VOYAGE_API_KEY
   • ANTHROPIC_API_KEY
3. Open the lab/ folder and run the notebooks in order.

The devcontainer automatically starts a local MongoDB Atlas-compatible instance and runs the seed script to populate the listings dataset.

Stack

• Runtime: TypeScript (tslab kernel in Jupyter)
• LLM: Claude Haiku via @langchain/anthropic
• Embeddings: VoyageAI voyage-4-large / voyage-4-lite
• Orchestration: LangGraph (@langchain/langgraph)
• Database: MongoDB Atlas local (mongodb/mongodb-atlas-local)