DEV Community

NILE GREEN
NILE GREEN

Posted on

The Game Genie for LLMs: Persistent Cognitive State for Your Agents

#agents #llm #opensource #showdev

thermomind‑continuity — Open Source (MIT)

LLMs are powerful, but they all share the same fatal flaw:

They forget everything after every message.

Every call resets the model. Every turn wipes identity. Every agent starts from zero.

Developers have been duct‑taping around this with:

  • Vector DBs
  • Prompt‑stuffing
  • RAG loops
  • Context juggling
  • Fragile memory hacks

None of these give you actual continuity. They just replay text.

So I built something different.

🎮 A Game Genie for Your AI

If you ever plugged a Game Genie into an NES cartridge, you already understand the idea.

You don't replace the game. You attach something that gives it abilities it never had.

thermomind-continuity does exactly that for LLMs.

npm install thermomind-continuity
pip install thermomind-continuity
Enter fullscreen mode Exit fullscreen mode

Wrap your existing OpenAI/Anthropic/DeepSeek client:

let openai = new OpenAI({ apiKey: process.env.OPENAI_API_KEY });
openai = tm.wrapOpenAI(openai);
Enter fullscreen mode Exit fullscreen mode

Now your agent has:

  • Persistent identity
  • Long‑term memory
  • Drift detection
  • Stability tracking
  • Surplus (growth energy)
  • Continuity guidance

No fine‑tuning. No GPUs. No rewriting your stack.

⚡ What It Actually Does

Every time your agent interacts, the backend computes:

Metric Meaning
Surplus How much "cognitive energy" it has to grow
Drift How far it's deviating from its own identity
Stability How coherent it is across sessions
Identity Its evolving fingerprint
Memory What it has retained over time

These metrics update automatically on every turn.

Your agent stops being a stateless loop. It becomes a persistent system.

🧩 Works With Any Model, Any Framework

Models Frameworks
GPT LangChain
Claude CrewAI
DeepSeek AutoGen
Open‑weights Raw API calls

If it speaks JSON, it works.

🏗️ Architecture Overview

  1. The SDK sends your session ID to the backend.
  2. The backend returns continuity hints based on:
    • Long‑term memory
    • Surplus / drift / stability
    • Identity fingerprint
    • Historical behavior
  3. These hints are injected into the system prompt before the LLM runs.

The LLM stays stateless. Your agent doesn't.

📈 Real Growth Over Time

This is what a real agent looks like after 200+ cycles:

Cycle  Surplus  Drift  Stability  Grade  Event
001    0.41     0.31   0.55       B      session_start
047    0.68     0.14   0.74       A      coherence_peak
134    0.74     0.09   0.88       A      identity_stable
200    0.81     0.07   0.91       A+     long_horizon_stable
Enter fullscreen mode Exit fullscreen mode

Same agent. 200 turns later. No resets.

🚀 Try It Yourself

Repo (MIT): https://github.com/nile-green-ai/thermomind-continuity

Start a session:

curl -X POST https://thermomind-production.up.railway.app/v1/sessions \
  -H "Authorization: Bearer test_public_key" \
  -d '{"external_id": "terminal-agent"}'
Enter fullscreen mode Exit fullscreen mode

Check its state:

curl -X GET https://thermomind-production.up.railway.app/v1/sessions/terminal-agent/state \
  -H "Authorization: Bearer test_public_key"
Enter fullscreen mode Exit fullscreen mode

🧠 Why This Matters

LLMs are incredible, but they're not agents. They're token predictors.

Agents need:

  • Memory
  • Identity
  • Continuity
  • Drift correction
  • Stability
  • Long‑horizon behavior

thermomind-continuity gives them that.

One line of code. Any model. Any framework.

Top comments (0)