Enhanced memory capability

[Ninot1Quyi]

Last Edited:2025-4-14 22:16

1. Modify Memory Storage Format

• [x] Change memory in memory.json from a single string to a list format with timestamps
• [x] Implement generating a summary every summary_chunk_size messages and storing it in the list
• [x] Ensure list elements follow the format: ["example summary 1", "example summary 2"]
• [x] Update memory storage to structured format with fields like example:

  "memories": [
    {
      "id": "base36_6bit",
      "created_at": "YYYY-MM-DD HH:MM:SS",
      "memory": "example memory description",
      "trigger_type": "auto/user_flagged",
      "message_timeframe": {
        "earliest": "YYYY-MM-DD HH:MM:SS",
        "latest": "YYYY-MM-DD HH:MM:SS"
      }
    }
  ]
  

  And history turns example:

  "turns": [
    {
      "role": "user/assistant",
      "content": "xxxx",
      "created_at": "2025-03-26 20:47:06"
    }
  ]
  

• [x] Test if the compressed history is neat and practical

2. Add New Command !reflectionToMem [num]

• [x] Add the !reflectionToMem [num] command to COMMAND
• [x] Implement default behavior to summarize the last 5 conversations and add to memory when no parameter is provided
• [x] Support custom conversation ranges when a parameter is given (e.g., !reflectionToMem 10 summarizes the last 10)
• [x] Add automatic extraction to summaries every summary_chunk_size messages
• [x] Test the command’s flexibility and correctness

3. Optimize Memory Saving and Summary Details

• [x] Adjust the saving_memory prompt in profiles/defaults/_default.json to be more reasonable
• [x] Adjust the ratio of summary to details to enable LLM to grasp information at different levels

4. Complete the CRUD of Memory

• [x] [Read] Memory retrieval: Integrate embedding vectors for memory relevance retrieval
  • Implemented semantic search with cosine similarity for context-aware memory retrieval
  • Created table format display with memory ID, type, timestamp and content columns
  • Modified display to hide timestamps for knowledge-type memories
• [x] [Create] Memory addition: Implemented structured memory creation with timestamps and type classification
  • Added LLM-driven memory operations via JSON commands (newMem)
• [x] [Update&Delete] Memory merging:
  • Added LLM-driven memory operations via JSON commands (mergeMem/noMerge)
  • Implemented memory type support (event/knowledge) with appropriate display formatting
  • Added robust error handling for partial merges when some memory IDs aren't found
  • Created detailed operation logging system for tracking all memory manipulations
• [x] Add a configuration option in settings.json to adjust retrieval count (default 10, -1 for all)

5. Code Validation

• [ ] Test whether the memory module is working properly
• [ ] Compare the new memory management method with the original method to check whether the performance is improved
• [ ] Count the change in the number of memories, and the change in the number of different kinds of memories

6. Code Optimization

• [ ] Optimize memory management prompt
• [ ] Check if the modified code is concise and meets project requirements
• [ ] Run full tests to ensure new features don’t affect existing functionality
• [ ] Discuss with Max or the team to confirm if the implementation meets expectations

Notes

• Mark tasks as completed with a check (- [x]) to track progress.
• Add any additional requirements or adjustments to this list as needed.

[Ninot1Quyi]

Last Edited:2025-4-20 14:59

I'm adding MCP support to supercharge the MindCraft community and unlock a ton of new possibilities for Andy! 🎉

Down the road, I’ll implement Memory as a pluggable MCP Server component, making it easy to swap in and out as needed.

[Ninot1Quyi]

Added the Memory tool from the official Model Context Protocol (MCP) example using MCP, enabling the storage of memories in a knowledge graph.

Please refer to the methods in this PR: https://github.com/kolbytn/mindcraft/pull/508#issue-3007100329

[Sweaterdog]

I would love to see this in Mindcraft-CE