pi-continuous-learning-v2: A Continuous Learning Evolution System for AI Programming Assistants

pi-continuous-learning-v2 is an independent extension package for pi-coding-agent, designed to address the pain point where AI programming assistants forget project-specific knowledge across sessions. It can automatically capture project 'intuitive' knowledge (such as coding standards, architectural conventions) from conversations, generate reusable skills and prompt templates, allowing the AI assistant to accumulate project experience like human developers and understand the current project better with more use.

With the popularity of AI-assisted programming tools today, a long-standing pain point is: during cross-session collaboration, project-specific tacit knowledge (coding standards, architectural conventions, common patterns, etc.) is lost, and the AI needs to re-learn the context in each new session. pi-continuous-learning-v2 is designed exactly to solve this problem, implementing the 'Continuous Learning v2' architecture to allow AI assistants to accumulate project experience.

The project adopts an independent extension package architecture, which does not interfere with the stability of the core system and can be iterated independently. Key technologies include:

1. Observation capture mechanism: Listens to pi extension events to collect interaction data (code modifications, file operations, etc.);
2. Intuition storage system: Two-layer architecture (project-level stored in <project>/.pi/continuous-learning-v2/, global-level stored in ~/.pi/agent/continuous-learning-v2/);
3. Background observation analyzer: Asynchronously analyzes accumulated data (triggered by default every 5 minutes or when the threshold of 20 entries is reached), without blocking user operations.

The project provides rich commands to manage the learning process:

• Intuition management: /instinct-status (check status), /instinct-export/import (import/export), /promote (promote pending intuition for review), /prune (clean up expired entries);
• Skill generation: /skill-create analyzes Git history to generate skill files (supports parameters like --commits depth, --output path, etc.);
• Evolution: /evolve --generate clusters intuitions into independent skills;
• Evaluation: /learn-eval checks learning quality to avoid duplication.

The system uses a confidence-level stratification strategy: high-confidence intuitions are directly written to active storage, while low-confidence ones are temporarily stored in instincts/pending/ for review. The deduplication mechanism avoids repeated learning (if a new observation is similar to existing knowledge, it is merged or skipped). Observation burst handling optimization: batch analysis reduces computational overhead. The TTL mechanism (default 30 days) manages the lifecycle of pending intuitions.

The system automatically detects Git project directories: when running inside a repository, the output belongs to the project; otherwise, it falls back to the global scope. It supports manual specification of the project root directory via PI_PROJECT_DIR and CLAUDE_PROJECT_DIR environment variables, compatible with complex development environments like CI/CD and containerization.

The first run generates a default configuration (~/.pi/agent/continuous-learning-v2/config.json), which includes parameters like observer switch and running interval. Model selection priority: session active model → configuration file observer.model → pi global default model, allowing configuration of a dedicated model for analysis tasks.

pi-continuous-learning-v2 evolves AI assistants from tools into collaborative partners that understand project context. It is suitable for long-term large-scale projects, team collaboration, and organizations that need consistent coding styles, helping to build exclusive AI knowledge bases. It inherits the ECC continuous learning concept but adapts to the pi native ecosystem, making it an important attempt to evolve AI-assisted programming towards 'memory-enabled'.