Gemini-Mem V3: The Self-Evolving Cognitive Engine

"Experience is the architect of the mind."

Gemini-Mem V3 is not a search index; it is a self-evolving cognitive substrate for the Gemini CLI. Inspired by the Evolver architecture and high-frequency agent workflows, V3 bridges the gap between ephemeral interactions and permanent engineering wisdom.

---

🧬 The "Evolver" Feedback Loop

V3 implements a four-stage evolutionary cycle that operates silently in the background:

1. Ingestion: Capture raw dialogue via zero-latency asynchronous hooks.
2. Extraction: Gemini 3 series models distill the "What" (Facts), the "Who" (Entities), and the "How" (SOPs).
3. Recognition: The engine detects recurring technical patterns (Concepts) across multiple sessions.
4. Distillation: Upon reaching a "Success Threshold," raw memories are crystallized into formal .md skills.

---

⚡ Hardcore Architectural Features

1. Procedural Memory: The SOP Engine

Traditional memory systems fail because they recall facts when you need processes. V3 extracts Standard Operating Procedures (SOPs) from your successes.

• Micro-Workflow Capture: Automatically identifies the exact shell commands, file patches, and config tweaks that solved a bug.
• Example Output:

  "sop": ["1. Bypass build isolation using --no-build-isolation", "2. Patch setup.py to hardcode sm_86", "3. Rebuild native modules"]

2. Knowledge Solidification (SkillDistiller)

The SkillDistiller is the heart of the engine's "Self-Learning" capability.

• Pattern Matching: Monitors the concepts_json field in the SQLite backend.
• The 2-Hit Rule: Once a complex technical concept (e.g., "Docker Optimization") is successfully validated twice, the Distiller triggers a synthesis event.
• Auto-Skill Creation: Generates a professional-grade SKILL.md in .gemini/skills/, creating a project-specific library of "Best Practices."

3. Asynchronous Nervous System

To prevent AI "brain fog" (UI lag), V3 decouples cognition from conversation.

• Task Offloading: The AfterAgent hook merely appends to a pending_tasks queue (WAL-mode SQLite).
• Background Worker: A dedicated process handles the heavy LLM lifting, ensuring your CLI remains responsive regardless of memory complexity.

4. Layered Semantic Retrieval

V3 uses a tiered approach to memory recall, ensuring that the most relevant "wisdom" is always injected:

• Tier 1: Entities: Matches specific symbols (functions, class names) currently in scope.
• Tier 2: Concepts: Injects architectural patterns relevant to the current task.
• Tier 3: SOPs: Provides actionable steps for the current problem based on historical wins.

---

🏗️ Technical Specifications

• Kernel: Optimized for Gemini 3 Series (leveraging high-token reasoning).
• Storage Layer:
  • SQLite FTS5: High-speed full-text indexing for titles, summaries, and SOPs.
  • JSON-Relation Mapping: Hierarchical storage of concepts and entities.
• Environment: Native Node.js / Bun compatibility with minimal dependency footprint.

---

🚀 Deployment & Evolution

Installation

gemini extension install .

Self-Learning in Action

1. Work: Solve a complex problem with the Gemini CLI.
2. Evolve: V3 background worker extracts the SOP and stores the concept.
3. Crystallize: Repeat a similar task. The SkillDistiller automatically creates a Skill file.
4. Recall: Next time you start a session, the relevant Skill is automatically injected into the context.

---

🛡️ Engineering Standards

• Local-First: Zero telemetry for your data. Your wisdom stays in your project.
• Anti-Leakage: Advanced regex redaction to ensure credentials never enter the long-term memory pool.

---

📜 License

GPL-3.0
Built to protect open-source innovation and prevent architectural plagiarism.

---

Designed for engineers who build things that last. Driven by an engine that never stops learning.