ResearchMind

A multi-agent AI research pipeline with cross-session memory

Type a topic. Three AIs collaborate. Get a polished research brief.
Come back tomorrow — it remembers everything.

Live Demo · Quick Start · How It Works · Architecture

---

Overview

ResearchMind is a multi-model AI pipeline built on the Backboard SDK. It decomposes any research topic into focused subtasks, processes each independently, and synthesizes the findings into a structured research brief — while persisting what you have researched across sessions so it never covers the same ground twice.

Built for the Backboard Challenges Hackathon · May 1–22, 2026

---

Live Demo

🚀 researchmind-production.railway.app

---

How It Works

Input Topic
    │
    ▼
┌─────────────────────────────────────────────────────┐
│  Stage 1 — Planner          (Gemini 2.5 Flash)      │
│  Reads memory → generates 3–5 focused subtasks      │
│  Skips anything already researched in past sessions │
└──────────────────────┬──────────────────────────────┘
                       │  subtasks[]
                       ▼
┌─────────────────────────────────────────────────────┐
│  Stage 2 — Summarizer       (Gemini 2.5 Flash)      │
│  Processes each subtask independently               │
│  Extracts 5–8 key facts per subtask                 │
└──────────────────────┬──────────────────────────────┘
                       │  findings[]
                       ▼
┌─────────────────────────────────────────────────────┐
│  Stage 3 — Synthesizer      (Claude Haiku)          │
│  Reads all findings                                 │
│  Writes polished research brief with headers        │
└──────────────────────┬──────────────────────────────┘
                       │  brief
                       ▼
            Research Brief Output
                       │
                       ▼
         Saved to Planner Cross-Session Memory

---

Architecture

Agent	Model	Responsibility
Planner	Gemini 2.5 Flash	Decomposes topic into subtasks. Reads memory to avoid repeating past research.
Summarizer	Gemini 2.5 Flash	Processes each subtask independently. Returns 5–8 key facts per subtask.
Synthesizer	Claude Haiku	Synthesizes all findings into a structured, polished research brief.

Backboard Features Used

Feature	Usage
Multi-assistant architecture	3 permanent assistants with distinct system prompts and roles
Multi-model routing	Gemini Flash for speed at scale, Claude Haiku for synthesis quality
Persistent cross-session memory	Planner stores topic + subtasks + conclusions after every session
Cross-thread memory recall	New sessions automatically skip previously covered subtasks

---

Quick Start

1. Clone

git clone https://github.com/Amulya631/ResearchMind.git
cd ResearchMind

2. Install dependencies

pip install -r requirements.txt

3. Configure environment

Create a .env file in the project root:

BACKBOARD_API_KEY=your-backboard-api-key

4. Create assistants (run once — idempotent, safe to re-run)

python assistants.py

5. Run

python -m streamlit run app.py

Open http://localhost:8501

---

Project Structure

ResearchMind/
├── app.py            Streamlit UI — live stage tracking, memory panel, download
├── assistants.py     Creates / reuses the 3 Backboard assistants (idempotent)
├── pipeline.py       Core pipeline logic — Planner → Summarizer → Synthesizer
├── requirements.txt  Python dependencies
├── .env              API key — never committed
└── .gitignore

---

Requirements

backboard-sdk>=1.5.0
streamlit>=1.32.0
python-dotenv>=1.0.0
fpdf2>=2.7.0

---

The Cross-Session Memory Feature

This is what separates ResearchMind from a standard chatbot.

After every pipeline run, the Planner stores:

• The topic researched
• Every subtask covered
• Key conclusions from the brief

On the next run, the Planner reads this memory before generating subtasks — so it never repeats research. Close the app, restart your machine, come back days later — the context is still there.

---

License

MIT License — see LICENSE for full terms.

Copyright © 2026 Batila Amulya

---

Built with Backboard · Streamlit · Google Gemini · Anthropic Claude

Batila Amulya · GCP Cloud Engineer & Gen AI