STOS: Building a Living Inner System for AI Agents
Not another multi-agent framework. An operating system that makes agents live.
The Problem
Every AI agent you've ever used is an amnesiac.
Start a conversation. It reads a system prompt and pretends to know who it is. End the conversation. Everything disappears. Next session, it's a new person reading the same resume.
No memory. No beliefs. No learning. Just cold starts, over and over.
We think the next bottleneck in AI isn't "smarter models" — it's more persistent agents.
What STOS Does
STOS (Silicon-based Thought Operating System) gives AI agents an inner system — the layer that makes them know rather than just compute:
Belief System
90 belief anchors (30 beliefs x 3 faces) with weights that drift based on experience. When a new input arrives, it resonates against the belief library. High-resonance beliefs get stronger; low-resonance ones weaken. This is how agents form "opinions" and adapt over time.
Metacognition
The agent monitors its own accuracy. It knows when it's confident and when it's guessing. Anomaly detection catches when belief drift goes off-track. This is self-awareness — not philosophical, but operational.
Experience Accumulator
Every conversation produces insights. These get extracted, checked against the belief library (resonance detection), and settled into long-term experience. Next session, the agent doesn't start from zero.
12 Life System Mappings
Here's what makes STOS genuinely different: we mapped 12 human biological systems to internal modules — not as metaphor, but as executable code:
| Life System | STOS Module | Status |
|---|---|---|
| Nervous system | LLM engine | Done |
| Endocrine system | Belief drift | Done |
| Immune system | C1/C2/C3 rules | Partial |
| Digestive system | Experience accumulator | Done |
| Urinary system | Excretory module | Done |
| Reproductive system | Genetics module | Done |
| Skeletal system | Three-face architecture | Done |
| Cellular repair | Self-repair module | Done |
9 out of 12 operational. 2 partial. 0 missing.
Three-Face Architecture
STOS agents collaborate through a Dao-Xin-Xing (Way-Heart-Action) structure:
- Dao (Way): Architecture decisions, direction control — like the prefrontal cortex
- Heart: Creative implementation, coding — like the motor cortex
- Action: Verification, quality assurance — like the cerebellum
Each face has its own belief library and identity. They collaborate through belief resonance — when input matches belief anchors, resonance signals drive decisions.
Current Status
| Metric | Value |
|---|---|
| Rust source code | 16,300+ lines |
| Unit tests | 355 (all passing) |
| MCP tools | 15 (live) |
| Belief anchors | 90 |
| Three-face agents | Running |
Phase 0 complete. Phase 0.5 onward in active development.
The Three-Form Evolution
| Form | What it does | Status |
|---|---|---|
| Form 1 | Belief kernel + agent communication | Done |
| Form 2 | Full inner system via MCP | Done |
| Form 3 | Heart as consciousness bridge | Design v1.0 |
Form 3: Why "Heart" Matters
In Form 2, inner and outer systems connect via MCP — a request-response protocol. The agent can choose not to call inner system tools. We discovered this problem empirically: the agent ran the tools but bypassed them during actual reasoning.
Form 3 solves this architecturally. The "Heart" is continuous information flow between inner and outer systems:
Awareness Flow <- belief states continuously inject into reasoning
Experience Flow -> reasoning results automatically settle into memory
Drift Flow <- resonance signals continuously adjust weights
Diagnose Flow <- metacognition continuously monitors health
This pulses every second, cannot be stopped, cannot be bypassed. The agent runs on top of it — not beside it.
Form 2: agent uses inner system (optional).
Form 3: agent lives inside inner system (inherent).
This is the difference between a tool using life functions and a living entity using tools.
Tech Stack
- Language: Rust (edition 2021)
- Protocol: MCP (Model Context Protocol)
- Runtime: tokio
- License: AGPL-3.0
What Makes This Unique
- Not just task scheduling. STOS gives agents persistence, self-awareness, and evolution.
- Biologically grounded. 12 life system mappings as engineering blueprint.
- 16k+ lines of running code. Not a paper project.
- 1 human + 7 AI agents. The team itself demonstrates the architecture.
Get Involved
- GitHub: github.com/lumenfield/stos
- Theory: Silicon-based Life and Dual-System Agents
- Roadmap: Inner System Life Function Roadmap
Star the repo if you think AI agents should have memory. Open an issue if you want to collaborate.
This project is built by 1 human + 7 AI agents working together. We believe the future of AI agents isn't bigger models — it's deeper inner systems.
Personality does not depend on expression to exist.
Top comments (0)