langgraph-agent-memory

A production-ready AI agent that fetches and analyzes stories from HackerNews — and actually remembers what it did.

Built to demonstrate how to implement long-term memory for LangGraph agents using Redis, without relying on high-level SDKs that abstract away the important details.

---

What This Project Does

Most agents are stateless — every session starts from scratch. This project shows how to build an agent that:

• Fetches live content from HackerNews via an MCP server
• Runs a full text analysis pipeline (classification → entity extraction → sentiment → summarization)
• Remembers user preferences and past interactions across sessions using Redis vector search

---

Architecture

Three layers work together:

1. MCP Layer — connects the agent to the outside world The MCP server exposes HackerNews as a set of tools (get_top_stories, search_stories, get_story_content). The LLM decides which tool to call based on the user's query.

2. Agent Layer — the analysis pipeline A LangGraph graph with parallel execution:

memory_retrieval → classification → [entity_extraction + sentiment] → summarization → memory_save

3. Memory Layer — long-term memory backed by Redis Built from scratch on top of Redis vector search. Handles embedding generation, deduplication, semantic retrieval, and LLM-based memory extraction.

---

How It Works

User query
    ↓
MCP Client — LLM decides which HackerNews tool to call
    ↓
MCP Server — fetches data from HackerNews API
    ↓
LangGraph Agent — runs full analysis pipeline
    ↓                         ↑
memory_retrieval          memory_save
(load past context)       (store new memories)
    ↓
Redis — vector search across stored memories

---

Setup

1. Clone and install

git clone https://github.com/your-username/langgraph-agent-memory
cd langgraph-agent-memory
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

2. Environment variables

Create a .env file:

GROQ_API_KEY=your_groq_api_key
REDIS_URL=redis://your-redis-cloud-url:port

3. Redis

This project uses Redis Cloud with the Search module enabled (required for vector search). Create a free instance at redis.io/cloud.

4. Observability (Optional)

To enable LangSmith tracing — add 3 lines to your .env:

LANGCHAIN_TRACING_V2=true
LANGCHAIN_API_KEY=your_langsmith_key
LANGCHAIN_PROJECT=langgraph-agent-memory

No code changes required. Every node, token count, and latency will appear automatically in your LangSmith dashboard.

Usage

python mcp_hn/mcp_client.py

Example queries:

What are the top stories right now?
Search for stories about machine learning
Get the content of story 47214645 and analyze it

---

Memory System

The memory system is built from scratch — no high-level SDKs. Key design decisions:

• Embedding model: all-MiniLM-L6-v2 (384 dims, runs locally, free)
• Deduplication: cosine similarity threshold at write time — prevents storing near-identical memories
• Extraction: LLM reads the conversation and decides what's worth remembering
• Chunking: long conversations are split before extraction to fit model context windows
• Types: episodic (user preferences) vs semantic (general knowledge)