Dungeon Explorer (Grimshaw) — Toy Agent Environment

This repo is a simple dungeon-explorer environment built to demonstrate an AI agent (“Grimshaw”) using:

• Scratchpad memory (short-term working notes)
• Fading memory (older facts decay / get summarized)
• Visual input of surroundings + resources
• Optional help from an external human “Oracle” agent

The goal is straightforward: survive long enough to reach the exit.

This is intentionally a toy example of a “self-owned business” pattern: an agent can pursue a goal, request human help when needed, and rate/compensate human assistance. The scaffolding here is meant to be reusable for real-world tasks.

---

Why this exists

Long-horizon tasks are hard for agents because they require:

• remembering what mattered,
• adapting to changing conditions,
• and knowing when to ask for help.

This toy environment lets you test those behaviors cheaply before attempting real-world workflows.

It currently uses Ollama with a local model (e.g., GPT-OSS 20B “thinking”). More advanced frontier models may do better. Better memory systems may also help with long-horizon tasks (knowledge graphs, vector DBs, fine-tuning smaller specialist agents, etc.).

---

Motivational dynamics (artificial “emotions” + resources)

The dungeon includes dynamic resources (energy/health/oil/etc.). On top of that, Grimshaw has lightweight artificial “emotions” (pressure/urgency) to inject motivation—not as “feelings,” but as a control mechanism that helps it keep moving toward the exit instead of stalling.

Observations so far

• Grimshaw tends to follow the Oracle more when the Oracle explains why the advice fits Grimshaw’s own state/constraints.
• Under extreme pressure (low resources), Grimshaw may hallucinate “boxes” that could save it. This is weird but useful to study: stress + incomplete grounding can create “story-like” survival behavior.

---

Repo notes (important fixes)

1) index.html must be in templates/

If you’re using Flask (or a Flask-like layout), index.html must live in folder templates.

Repo notes (important fixes)

1) index.html must be in templates/

If you’re using Flask (or a Flask-like layout), index.html must live here:

templates/index.html

Flask’s render_template("index.html") looks in the templates/ directory by default.

2) OAuth / email / Graph remnants

You said you removed email + OAuth Graph mailing, but some configs may still be present. If anything still references OAuth credentials or email sending:

• Disable those code paths, or
• Update the credentials + redirect URLs to your current setup.

Practical tip: search the repo for these strings and remove/disable anything you don’t use:

• CLIENT_ID, CLIENT_SECRET, TENANT_ID, REDIRECT_URI
• graph.microsoft.com, Authorization: Bearer
• send_mail, smtp, OAUTH, MSAL
• change xxx@gmail.com to your own emails

Requirements

• Python 3.10+ (3.11 recommended)
• Ollama installed and running
• The Ollama model you want to use pulled locally (example below) flask (web app + API)

Files use: (really standard stuff) -requests (Microsoft Graph token + sendMail) if EMAIL_ENABLED -urllib.request (Ollama /api/generate) -re (JSON extraction, action/position fallback parsing) -socket (local IP detection for dashboard link) -standard libs: json, time, threading, random, dataclasses, typing, enum, collections (Counter/Deque)

Quick start (how to run)

1) Start Ollama

Make sure Ollama is running:

example (adjust to your actual local model id):

ollama pull gpt-oss20b-thinking