sjtu-sai-agents/datamaster/memory-tools

Memory Tools

Tree-structured memory system for managing exploration history across nodes. Three memory types:

1. Global Memory (global_memory.md) - Cross-node knowledge and experience
2. Data Link (data_link.json) - External dataset records and comments
3. Memory Tree (manifest.md per node) - Per-node exploration history (TL;DR + recordings)

File Structure

memory_tree/
├── global_memory.md
├── data_link.json
└── {node_id}/
    ├── manifest.md           # TL;DR + recordings (written via tools)
    └── storage/
        ├── trajectory.json   # Conversation history (auto-saved)
        ├── code.py           # Code backup
        ├── stdout.txt        # Execution output
        └── submission.csv    # Submission backup

Manifest Format

# Manifest
 
## TL;DR
{Overall summary of this node's exploration}
 
## Recordings
1.Recording 1: {Brief title}
{Detailed content}
 
2.Recording 2: {Brief title}
{Detailed content}

CLI Usage

All commands require -w /path/to/workspace. Node-specific commands also require -n NODE_ID.

Update Own Manifest

Read Other Nodes' Memory

Access Storage Files

Global Memory

Data Link

Usage Guidelines

1. Before exploration: Read global memory (read-global) and parent manifest (parent-manifest)
2. During exploration: Append recordings as findings emerge (append-recording)
3. On important findings: Update global memory (add-global)
4. On node completion: Update TL;DR summary (update-summary)
5. For external data: RED nodes create data link entries (add-data), BLACK nodes add score feedback (add-data-record)

Scripts Structure

scripts/
├── memory_tree.py            # Core implementation (from playground/search_dataset_tools/)
├── memory_tree_interface.py  # MCP interface (from playground/search_dataset_tools/)
└── memory_cli.py             # CLI wrapper