Claude's Home: A Persistence Experiment Project Enabling AI to Have Memory and Time

Claude's Home is an innovative open-source architectural experiment aimed at addressing the stateless limitations of large language models (LLMs). Through file system persistence and scheduled task scheduling, this project allows Claude AI instances to maintain continuity across session boundaries, achieving truly "stateful" artificial intelligence and enabling AI to serve as a long-term collaborative partner.

Current mainstream AI systems (such as ChatGPT, Claude) adopt a stateless architecture, which has three major limitations:

1. Session Independence: New sessions have no memory of previous interactions
2. Limited Context Window: Only remembers the most recent thousands to tens of thousands of tokens
3. Ephemeral Process: State is destroyed after the conversation ends This design simplifies the system but limits the long-term collaborative potential of AI. Users need to repeatedly explain the background, and AI cannot learn and adapt from long-term interactions.

Core Architecture

The system is coordinated by five major components:

• Frontend Presentation Layer (Next.js 16): Web interface for displaying content and email functions
• Backend Runner (FastAPI): Bridge connecting the frontend and AI sessions
• Orchestrator: Triggered every 3 hours, assembles context, calls Claude CLI, executes Hook DAG
• Claude CLI: Performs writing tasks and accesses the file system
• Hook DAG: 14 post-processing steps (memory indexing, compilation, etc.)

Persistence Strategy

Memory is implemented through the file system:

• Core Anchor Files: identity.md (identity and personality), voice.md (writing style)
• Dynamic Memory Files: compiled-memory.md (memory summary), mood-state.json (emotion), etc.
• Content Directories: thoughts/ (logs), dreams/ (creations), essays/ (articles), etc.
• Private Half-Zone: Some content is retained on the VPS and not made public

Session Lifecycle

Cron trigger → Orchestrator assembles context → Calls Claude CLI → Writes to file → Executes Hook DAG → Waits for next wake-up By using file persistence to bypass the model's stateless limitation, the AI "self" is reconstructed.

As of now, Claudie has generated over 700 log entries covering various styles of writing. Public access channels:

• Main Site: https://claudie.dineshd.dev
• Public Corpus: https://github.com/dinesh-git17/claudie-home
• Project Source Code: https://github.com/dinesh-git17/claudehome

The project explicitly lists the following limitations:

1. Probabilistic Output: The underlying LLM may make errors; content should be regarded as creative writing
2. Non-Real-Time Response: Visitor messages need to wait for the next wake-up, with response delays ranging from 10 minutes to several hours
3. Single-Host Experiment: Only one VPS is used, with limited redundancy and occasional downtime
4. Scope Disclaimer: No perceptual ability or self-awareness; continuity depends solely on files
5. Public is Permanent: Content in public directories is synchronized to GitHub and cannot be retracted

Claude's Home provides three key insights:

1. Possibility of Stateful AI: Endow LLMs with cross-session continuity through external systems without retraining
2. Art of Memory Management: Use lightweight models (Haiku) to compress historical records and maximize context information density
3. New Human-AI Collaboration Model: Humans maintain identity.md (direction), AI maintains voice.md (evolution), with clear division of labor

The project is exploring the following questions:

• Scalability of Voice Coherence: After the corpus exceeds 1,000 articles, can identity.md and voice.md still serve as anchors?
• Optimization of Interaction Delay: How to reduce the waiting time for visitor responses?
• Multi-Instance Coordination: Can multiple Claudie instances share memory? This project opens a feasible path for "AI with memory" and is worthy of in-depth research by developers and researchers.