Building an AI Research Agent with Persistent Memory

TL;DR

The Quercle Research Agent is an AI assistant that remembers what it learns. Unlike standard chatbots that forget everything between sessions, this agent builds knowledge over time using Quercle for web research, OpenRouter for LLM access, and MongoDB for persistent memory.

Links:

• GitHub Repository

The Problem with AI Research

You're researching AI agent frameworks. You spend 30 minutes with ChatGPT, learn a lot, then move on. Weeks later, you want to continue - but what exactly did you learn? Which sources did you use? What's changed since then?

Standard chatbots aren't designed for ongoing research. Even with memory features, they store general facts about you - not structured research findings with sources, timestamps, and topics you can query later.

The Quercle Research Agent solves this by maintaining a dedicated research memory you can search and build on over time.

How It Works

When you ask a question, the agent:

1. Checks memory first - Do I already know this?
2. Searches if needed - Uses Quercle to find current information
3. Stores findings - Saves key facts to MongoDB for future sessions
4. Cites sources - Every answer includes source links

Both your research memories and conversation history are stored in MongoDB - not browser storage. Your data persists across devices and won't disappear if you clear your browser.

Example: Building Knowledge Over Time

Session 1: Initial Research

You: What are the main approaches to AI agent memory?

Agent: [recall] No previous research found
       [search] "AI agent memory architectures 2025"
       [fetch] langchain.com/docs/memory
       [remember] Stored 6 findings about agent memory

Report: There are three main approaches...
        1. Buffer memory (recent context)
        2. Vector memory (semantic search)
        3. Entity memory (knowledge graphs)

Session 2: Follow-up (days later)

You: How does LangChain implement vector memory?

Agent: [recall] Found 2 memories about LangChain memory...
       [search] "LangChain vector memory implementation" (for details)
       [remember] Updated findings with implementation specifics

Report: Based on my previous research and new findings...

Session 3: Checking for updates

You: Any new developments in AI agent memory since my last research?

Agent: [recall] Last researched: 2 weeks ago
       [search] "AI agent memory latest developments"

Report: Since your last research, two notable developments...

Who Is This For?

Individual researchers, analysts, and curious minds who:

• Track evolving topics over weeks/months
• Return to the same domains repeatedly
• Want to see exactly what sources the AI used
• Prefer building cumulative knowledge over one-shot answers

Not for:

• One-off questions (just use ChatGPT/Claude/Perplexity)
• Team collaboration (no multi-user support yet)
• Real-time data needs (memory is point-in-time snapshots)

Tech Stack

Agent Tools

Quick Start

git clone https://github.com/quercledev/quercle-research-agent.git
cd quercle-research-agent
bun install
cp .env.example .env
# Edit .env with your API keys
bun dev

Configuration

# .env
QUERCLE_API_KEY=qk_...           # Web search
MONGODB_URI=mongodb+srv://...     # Memory storage
OPENROUTER_API_KEY=sk-or-...      # LLM provider
MODEL=openai/gpt-5-nano           # Or any OpenRouter model that supports tool calls

Limitations

• Single user - No authentication or multi-user support
• No sharing - Can't share research or collaborate
• Memory grows - No automatic cleanup of old memories

Get Started

Clone the repo, add your API keys, and start building cumulative knowledge:

github.com/quercledev/quercle-research-agent

Have ideas for new features or want to contribute? Pull requests and suggestions are welcome on GitHub!