TarsTARS

Self-Modification

How Tars evolves its own capabilities through memory, skills, and extensions.

Overview

One of Tars’ defining features is its ability to modify and extend itself. Through the combination of writable memory, skill creation, and extension development, Tars forms a feedback loop that enhances its capabilities over time.

Memory: The Living Brain

Tars uses an Episodic Session architecture combined with a dedicated MCP Memory server (tars-memory). This is not a static configuration file—Tars actively reads and writes to this database during conversations.

Examples of self-modification:

  • Recording user preferences (code style, preferred tools) into core facts
  • Noting project context (repo structure, deployment targets)
  • Tracking operational state (ongoing tasks, blockers) inside daily log notes

Skill Generation

When Tars encounters a recurring task or a complex workflow, it can create a new skill:

  1. User asks Tars to do something complex
  2. Tars solves it and recognizes the pattern
  3. Tars writes a new SKILL.md in ~/.tars/.gemini/skills/
  4. Future invocations reference the skill for consistent execution

Extension Development

Tars can create entirely new MCP extensions to integrate with external systems or add complex tools:

  1. Tars identifies a need for a new tool capability.
  2. Uses the extension-builder skill to scaffold the project.
  3. Generates a plain JavaScript MCP server (ESM) to avoid compilation overhead.
  4. Creates a gemini-extension.json manifest with native ${extensionPath} tokens.
  5. Authorizes the extension in extension-enablement.json.
  6. The tools are available after a quick tars restart.

The Feedback Loop

User Request → AI Processes → Learns Pattern
      ↓                            ↓
  Responds                  Updates Memory Facts
                              Creates Skill
                            Builds Extension
      ↓                            ↓
  Next Request ← Enhanced Capabilities ←

This loop means Tars becomes more capable over time, tailored specifically to the user’s workflow and preferences.

Safety

  • Memory changes are transparent — it’s written to plain JSON and Markdown files
  • Skills are human-readable markdown documents
  • Extensions are TypeScript source code you can audit
  • All modifications happen in ~/.tars/, isolated from critical system files unless you explicitly grant permissions
Next Up Supervisor Engine →