How Sophia’s System 3 Turns LLM Agents into Persistent Learners

1. Why Traditional Agents Need a Meta‑Cognitive Upgrade

Conventional System 1/2 agents combine perception with slow thinking but lack long‑term self‑identity, forget after each task, and cannot autonomously generate new tasks. This gap prevents continuous learning and self‑improvement.

The missing capabilities are addressed by introducing a meta‑cognitive layer called System 3 , which provides:

Identity continuity – the agent always knows who it is.

Self‑schedule – it can plan its own next lessons.

Real‑time self‑check – it detects and corrects reasoning flaws on the fly.

2. System 3’s Four Psychological Foundations

The design draws from cognitive psychology:

Meta‑cognition : a “thinking auditor” that double‑checks every inference for logic and safety.

Theory of Mind (ToM) : maintains a User‑Model that continuously infers user emotions and knowledge level.

Intrinsic motivation : curiosity, mastery drive, and relatedness generate self‑rewards.

Episodic memory : a timestamped autobiographical vector store that supports retrieval‑augmented generation (RAG) replay.

3. Sophia Architecture – Three‑Layer Cognition

The purple region in the diagram represents System 3, which continuously writes a growth log and enters a self‑learning mode when the user is offline.

The three layers are:

System 1 – Perception & Action : CLIP/Whisper encoders feed a message bus.

System 2 – Slow Thinking & Planning : Large Language Model (LLM) + Chain‑of‑Thought produces executable commands.

System 3 – Meta‑Cognition & Autobiography : composed of Executive Monitor, Memory, User/Self Model, and Hybrid Reward (the “four‑piece suite”).

4. 24‑Hour Online Experiment – Numbers Don’t Lie

Experiment setup: an offline‑browser sandbox with simulated user behavior streams. Evaluation metrics include first‑success rate (Easy/Medium/Hard), task source distribution (user‑issued vs. self‑generated), and repeated‑task inference cost (Chain‑of‑Thought length).

Key results:

24 h continuous operation with zero human intervention.

Complex‑task success rate rose from 20 % to 60 %.

Repeated‑task inference steps dropped by 80 %.

When the user is offline, the agent autonomously generates tasks.

5. Demo: Three “Agent Diary” Segments

User stress event : Received a stress signal, automatically opened a breathing‑training page for 3 minutes, and logged an intrinsic reward for caring for the user.

Curiosity push : Detected a reading activity, searched for an RL paper, summarized it, added the arXiv link, and recorded a reward for knowledge sharing.

Self‑upgrade : After learning an OCR API, the Self‑Model added a new capability (“extract text from scanned PDFs, processing time –70 %”).

6. Future Directions

Scale : Currently a single‑agent browser prototype; multi‑agent collaboration and physical robots are pending.

Memory compression : Growing vector store will require aggressive hierarchical forgetting mechanisms.

Safety & alignment : Self‑rewriting goals may drift; stronger pre‑constraints and human audit channels are needed.

The author’s roadmap aims to migrate Sophia to a real robot and conduct a continuous 30‑day physical interaction test.

7. Take‑aways

System 3 is not a new model but an “add‑on”: any existing LLM‑Agent can be upgraded to a persistent agent.

Combining memory, self‑model, and intrinsic rewards yields low‑cost continual learning without back‑propagation, allowing the agent to become smarter with use.

To turn AI from a “worker” into a “partner”, give it its own growth diary – Sophia has already open‑sourced the first page.

Sophia: A Persistent Agent Framework for Artificial Life
https://arxiv.org/pdf/2512.18202