Skip to content
Finalist

Voyager-VRAM

Voyager-VRAM is a long-horizon workplace simulator where LLM agents must navigate a realistic office environment ( emails, Slack, shared drive, spreadsheets, calendars ) using 31 tools to complete multi-step project mana

Code Demo Video Cerebral Valley

Core trick

Simulated office with 31 tools and three memory probes tests whether agents retain episodic, semantic, and working state across long tasks.

Tags

agentsproductivity

Description

Voyager-VRAM is a long-horizon workplace simulator where LLM agents must navigate a realistic office environment ( emails, Slack, shared drive, spreadsheets, calendars ) using 31 tools to complete multi-step project management tasks like preparing client briefs, resolving calendar conflicts, and reconciling budgets. On top of the environment we layer a Voyager-inspired learning system (skill library, working memory, episodic memory) with three types of memory probes (episodic, semantic, working), directly targeting the memory consolidation problem that current research (MEM1, Memory-R1) is trying to solve.
himalayadua/VRAM★ 20 forksCreated Mar 7, 2026 · last push Mar 8, 2026

README excerpt

<h1 align="center">🧠 Voyager-VRAM</h1>
<h3 align="center"><em>A Workplace Simulator for Training Memory-Aware Agents</em></h3>

<p align="center">
  <a href="#the-voyager-architecture"><img src="https://img.shields.io/badge/Architecture-Voyager--Lite-blue?style=for-the-badge" alt="Architecture"></a>
  <a href="#system-details"><img src="https://img.shields.io/badge/Model-Qwen2.5--1.5B-orange?style=for-the-badge" alt="Model"></a>
  <a href="#training--results"><img src="https://img.shields.io/badge/Training-Expert_Iteration-green?style=for-the-badge" alt="Training"></a>
</p>

---

## The Problem

Imagine you're managing a 6-week software project. You get 100 Slacks, 50 emails, meeting notes, spreadsheets, calendar invites — information comes at you constantly. Three weeks in, the client asks:

> *"When did we agree to cut the mobile feature?"*

…and you have to remember a decision buried in a Slack thread from week one.

The best project managers have **extraordinary working memory**. They remember who said what, which tasks block which, and what changed since last week.

**Can we train an LLM to develop that same skill?**

That's what we built.

<p align="center">
  <img src="images/HF-space.png" alt="VRAM-PM Hugging Face Space — Live Environment Demo" width="85%" />
  <br />
  <em>VRAM-PM: Live environment on Hugging Face Spaces</em>
</p>

---

## System Details

| Component | Detail |
|---|---|
| **Model** | Qwen2.5-1.5B-Instruct (4-bit quantized via Unsloth) |
| **LoRA Ran

Team