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.

3. Memory Retrieval

Ensure session-context-loader is configured to read this memory file at the start of every new chat session.

4. Cross-Project Episodic Memory (Global Context)

• Continuous Capture: For complex architectural decisions or bug fixes that span beyond a single project, extract the root cause and solution into a global knowledge graph (e.g., ~/.gemini/knowledge or equivalent <appDataDir>/knowledge KI system).
• Contextual Retrieval: On new sessions, search the global context directory for similar past issues before diving into complex bugs.
• Synthesis & Injection: Inject the historical context directly into the current execution loop, stating explicitly: "I remember we solved this architectural issue in Project X using technique Y. Applying the same proven pattern here."

---

Integration with Other Skills (MANDATORY)

• session-context-loader — Essential for loading the LEARNING_GRAPH.md when a new session starts.
• project-context-mapper — The context map tracks where things are; the memory graph tracks how and why things are built.

Referenced By Orchestrators (MANDATORY)

• brainstorming — Add to "Discovery & Audit".
• zero-to-prod-orchestrator — Phase 1 (Discovery & Architecture).

---

Bahasa Indonesia

Integrasi Orkestrasi

Terhubung dan mengorkestrasi skill domain yang relevan seperti brainstorming, zero-to-prod-orchestrator, dan project-context-mapper untuk memastikan eksekusi yang kohesif.

Deskripsi

Claude melupakan segalanya saat sesi chat baru dimulai. Skill ini memberi Anda "kekuatan super" berupa Ingatan Jangka Panjang yang berevolusi sendiri. Agen secara aktif mencatat preferensi koding pengguna, pelajaran dari bug yang sulit dipecahkan, dan keputusan arsitektur ke dalam file persisten LEARNING_GRAPH.md.

Kondisi Pemicu

• Saat pengguna mengoreksi gaya koding Anda (misal: "Gunakan bahasa Indonesia baku", "Jangan pakai var").
• Setelah Anda berhasil memecahkan bug yang menghabiskan waktu lama.
• Saat pengguna berkata "Ingat ini untuk proyek ke depannya."

Panduan Singkat

• Jangan Hanya Berjanji: Jika pengguna mengoreksi Anda, jangan sekadar menjawab "Baik, saya akan ingat." LLM tidak punya ingatan bawaan antar-sesi. Anda wajib menulis koreksi tersebut ke dalam file LEARNING_GRAPH.md di folder .agents/ atau direktori root.
• Kategorisasi Ingatan: Pisahkan ingatan menjadi: Preferensi Gaya (Style), Pelajaran Berharga (Lessons Learned), dan Keputusan Arsitektur (Decisions).
• Semakin Lama Semakin Pintar: Dengan membaca file ini di awal setiap percakapan, Anda tidak akan pernah mengulangi kesalahan yang sama dua kali.

Memori Episodik Lintas-Proyek (Konteks Global)

• Penangkapan Berkelanjutan: Untuk bug yang rumit dan pola arsitektur tingkat lanjut, ekstrak solusi ke dalam grafik pengetahuan global (misal: ~/.gemini/knowledge).
• Pencarian Kontekstual: Sebelum memecahkan masalah kompleks baru, cari direktori konteks global untuk kasus serupa.
• Sintesis & Injeksi: Secara proaktif sebutkan: "Saya ingat kita menyelesaikan ini di Proyek X dengan cara Y, saya akan menerapkannya di sini."