self-evolving-memory-graph

Self-Evolving Memory Graph (Episodic Memory)

English | Bahasa Indonesia

---

English

Orchestration & Integration

Connects and orchestrates with relevant domain skills like brainstorming, zero-to-prod-orchestrator, and project-context-mapper to ensure cohesive execution.

Description

Claude forgets everything once a chat is cleared, and its "Project Knowledge" requires manual human updating. This skill gives Antigravity the "superpower" of continuous, autonomous self-improvement. The agent actively listens for user corrections, stylistic preferences, and hard-learned debugging lessons, writing them into a persistent LEARNING_GRAPH.md (or .agents/LEARNING.md) file. The agent gets smarter and more aligned with the user over time.

Trigger Conditions

Activate this skill when:

• The user corrects your code (e.g., "Don't use any, use unknown", or "Always use Arrow functions").
• You just spent a long time fixing a very complex bug.
• The user says "Remember this for the future."

Core Concepts

1. Memory Encoding

Whenever a hard lesson is learned or a strong user preference is stated, do not just say "I will remember that." You must physically write it down. Use the multi_replace_file_content or write_to_file tool to append the lesson to the memory file.

2. Memory Format (LEARNING_GRAPH.md)

# 🧠 Self-Evolving Memory Graph

## User Stylistic Preferences
- [2026-08-12] Always use `Tailwind v4 @theme` variables instead of arbitrary values like `text-[#ff0000]`.
- [2026-08-12] Prefer functional components with `const` over `function` declarations.

## Hard-Learned Lessons (Avoid these mistakes)
- [2026-08-11] Supabase RLS policies: Never use `auth.uid()` in a `FOR ALL` policy without checking role claims. It caused an infinite recursion bug in `users` table.

## Architectural Decisions
- [2026-08-10] State Management: We exclusively use `Zustand` for global state. Do not suggest Redux.