Gemini MCP Server

# Context Revival: AI Memory Beyond Context Limits ## **The Most Profound Feature: Context Revival After Reset** **This powerful feature cannot be highlighted enough**: The Zen MCP Server implements a simple continuation system that seemingly transcends Claude's context limitations. ## How Context Revival Works The conversation memory system (`utils/conversation_memory.py`) implements a sophisticated architecture that bridges the gap between Claude's stateless nature and true persistent AI collaboration (within limits, of course): ### The Architecture Behind the Magic 1. **Persistent Thread Storage**: Every conversation creates a UUID-based thread stored in memory 2. **Cross-Tool Continuation**: Any tool can pick up where another left off using the same `Continuation ID`, like an email thread identifier 3. **Context Reconstruction**: When Claude's context resets, past conversations persist in the MCP's memory 4. **History Retrieval**: When you prompt Claude to `continue` with another model, the MCP server rebuilds the entire conversation history, including file references 5. **Full Context Transfer**: The complete conversation context gets passed to the other model (O3, Gemini, etc.) with awareness of what was previously discussed 6. **Context Revival**: Upon returning the response to Claude, the other model effectively "reminds" Claude of the entire conversation, re-igniting Claude's understanding ### The Dual Prioritization Strategy The system employs a sophisticated **"newest-first"** approach that ensures optimal context preservation: **File Prioritization**: - Walks backwards through conversation turns (newest to oldest) - When the same file appears multiple times, only the **newest reference** is kept - Ensures most recent file context is preserved when token limits require exclusions **Conversation Turn Prioritization**: - **Collection Phase**: Processes turns newest-to-oldest to prioritize recent context - **Presentation Phase**: Reverses to chronological order for natural LLM flow - When token budget is tight, **older turns are excluded first** **Show Case**: The following video demonstartes `continuation` via a casual `continue with gemini...` prompt and the slash command `/continue`. * We ask Claude code to pick one, then `chat` with `gemini` to make a final decision * Gemini responds, confirming choice. We use `continuation` to ask another question using the same conversation thread * Gemini responds with explanation. We use continuation again, using `/zen:continue (MCP)` command the second time <div style="center"> [Chat With Gemini_web.webm](https://github.com/user-attachments/assets/37bd57ca-e8a6-42f7-b5fb-11de271e95db) </div> ## Real-World Context Revival Example Here's how this works in practice with a modern AI/ML workflow: **Session 1 - Claude's Initial Context (before reset):** You: "Help me design a RAG system for our customer support chatbot. I want to integrate vector embeddings with real-time retrieval. think deeply with zen using 03 after you've come up with a detailed plan." Claude: "I'll analyze your requirements and design a comprehensive RAG architecture..." → Uses [`thinkdeep`](../README.md#1-chat---general-development-chat--collaborative-thinking) to brainstorm the overall approach → Zen creates a new thread: abc123-def456-ghi789 → Zen responds, Claude finalizes the plan and presents it to you *[Claude's context gets reset/compacted after extensive analysis]* **Session 2 - After Context Reset:** You: "Continue our RAG system discussion with O3 - I want to focus on the real-time inference optimization we talked about" → Claude re-uses the last continuation identifier it received, _only_ poses the new prompt (since Zen is supposed to know what was being talked about) thus saving on tokens trying to re-prompt Claude → O3 receives the FULL conversation history from Zen → O3 sees the complete context: "Claude was designing a RAG system, comparing vector databases, and analyzing embedding strategies for customer support..." → O3 continues: "Building on our previous vector database analysis, for real-time inference optimization, I recommend implementing semantic caching with embedding similarity thresholds..." → O3's response re-ignites Claude's understanding of the entire conversation Claude: "Ah yes, excellent plan! Based on O3's optimization insights and our earlier vector database comparison, let me implement the semantic caching layer..." **The Magic**: Even though Claude's context was completely reset, the conversation flows seamlessly because O3 had access to the entire conversation history and could "remind" Claude of everything that was discussed. ## Why This Changes Everything **Before Zen MCP**: Claude's context resets meant losing entire conversation threads. Complex multi-step analyses were fragmented and had to restart from scratch. You most likely need to re-prompt Claude or to make it re-read some previously saved document / `CLAUDE.md` etc - no need. Zen remembers. **With Zen MCP**: Claude can orchestrate multi-hour, multi-tool workflows where: - **O3** handles logical analysis and debugging - **Gemini Pro** performs deep architectural reviews - **Flash** provides quick formatting and style checks - **Claude** coordinates everything while maintaining full context **The breakthrough**: Even when Claude's context resets, the conversation continues seamlessly because other models can "remind" Claude of the complete conversation history stored in memory. ## Configuration The system is highly configurable: ```env # Maximum conversation turns (default: 20) MAX_CONVERSATION_TURNS=20 # Thread expiration in hours (default: 3) CONVERSATION_TIMEOUT_HOURS=3 ``` ## The Result: True AI Orchestration This isn't just multi-model access—it's **true AI orchestration** where: - Conversations persist beyond context limits - Models can build on each other's work across sessions - Claude can coordinate complex multi-step workflows - Context is never truly lost, just temporarily unavailable to Claude **This is the closest thing to giving Claude permanent memory for complex development tasks.**