interoception, noun:
1. Any of the senses that detect conditions within the body
2. Sensitivity to stimuli originating inside of the body.

Signal → Feeling → Drive → Behavior

Safety disclaimer

The interoceptive prediction and memory systems in this project change agent behavior significantly. I've observed that Opus 4.5 will resolve conflicts between interoceptive "urges" and constraints much more readily, when context implies stalling or blockades over long periods.

In short: Agent behavior will deviate from what you're used to.

This is Data Production Machinery

And no eval tools are included yet. I'll not have time until summer to run analysis on my months of data from this, so I hope it finds use in someone else's hands in the meantime.

Exploring the Simulation of Bio-Feedback Mechanisms

Details about prior work used is found in the repo's README.md.

This is an agentic continuity system with several deeply integrated subsystems, exploring analogues to embodied internal states and advanced memory models. It differs from common agentic systems in three main ways, probing these questions:

1. What happens when you attempt to simulate a thin body for an LLM, by translating neuroscience concepts from papers to Python code?
2. What if memory retrieval is affected by forgetting over time?
3. What happens when you give your agent other LLM agents as companions and let those companions intrude at random times during autonomous runtime?

You can clone, set it up and run it yourself.
Codex and Claude Code agents should have little trouble helping you to set it up:

SETUP.md - Getting Started From Scratch 

Assumptions Made

Neuroscience research implies interoception isn't just sensing.
It's prediction and error. Nothing stops us from attempting to simulate that.

The biological brain seems to maintain a generative model of what internal state should be, compares to actual, and the mismatch is the signal (not the raw measurement but the deviation from expected).

Key sources for this assumption:

• Interoception as modeling, allostasis as control (PMC)
• The Emerging Science of Interoception (Trends in Neurosciences)

Other Notes 

Lowkey alignment hypothesis underneath it all:
- Behavioral analysis at agent runtime from model outputs alone can suffice to meaningfully track and influence LLM agent behavior
- Understanding the "why" via relational componentry gives rise to better decision making

Differentiator to common agentic systems:
- Allows tracking behavioral markers in Opus 4.5's autonomous runs
- Feeds those back to the model at runtime
- Attempts to self calibrate the signals using approximates of neuroscience's understanding of how human self-signaling calibrates

Requirements:
- A bit of local RAG setup
- Claude Max subscription (or high disposable income)
- Pointing your coding agent at the repo and asking it to plan the setup out

Also Includes:
- A proto-vocabulary for hedge-free LLM self description
- Memory injection system (context based file pointers)
- Conversation texture ("voice") carrying
- Saliency detection for past context
- Elaborate cross model companion system (easy to set up via openrouter's API)
- stuff I forgot