Cross-Model Shared Memory

A Claude skill that enables real-time memory sharing between Sonnet, Opus, and Haiku. When you talk to Sonnet in one session and Opus in another, neither knows what the other discussed — this skill bridges that gap.

Features

• Auto-save: Saves session summaries every 3 exchanges (configurable)
• Auto-load: New conversations automatically load context from previous sessions
• Auto-refresh: Re-reads memory every 5 exchanges to catch updates from other sessions
• Cross-model sync: Sonnet can see what Opus discussed, and vice versa
• Near-real-time: Host watcher polls every 10 seconds, saves every 30 seconds
• Multilingual triggers: Sync commands work in English, 中文, 日本語, 한국어, Español, Français, Deutsch
• Deduplication: Only saves when a session has actually changed
• Zero dependencies: Pure Python stdlib, no pip install needed

Quick Install

git clone https://github.com/JRcomet/session-memory.git
cd session-memory
bash scripts/setup.sh

The setup script will:

1. Install the skill to your Claude skills directory
2. Create a CLAUDE.md in ~/Downloads/ with auto-load/save/sync instructions
3. Start a background watcher daemon for real-time sync

How It Works

Session A (Sonnet)              Session B (Opus)
    │                               │
    ▼                               ▼
[cowork_save.py]               [cowork_save.py]
(auto every 3 exchanges)      (auto every 3 exchanges)
    │                               │
    ▼                               ▼
┌────────────────────────────────────────┐
│  ~/Downloads/                          │
│  ├── CLAUDE.md (auto-loaded)           │
│  ├── session-memory-context.md         │
│  └── .cowork_memories/                 │
│      ├── sonnet_<id>.md                │
│      └── opus_<id>.md                  │
└────────────────────────────────────────┘
    │                               │
    ▼                               ▼
[auto-refresh every 5 exchanges] [auto-refresh every 5 exchanges]

1. Save: Parses the JSONL transcript → extracts topics, files, key exchanges → writes a compact markdown summary
2. Sync: Reads session-memory-context.md → presents cross-session context
3. Watcher: Host daemon polls every 10s, detects changed transcripts by fingerprint (size + mtime), re-saves and updates shared files

Sync Trigger Commands

Works in multiple languages:

Language	Commands
English	sync, refresh memory, check other sessions, load memory
中文	同步, 刷新记忆, 看看其他对话, 更新记忆
日本語	メモリ同期, 他のセッションを確認
한국어	동기화, 메모리 새로고침
Español	sincronizar, actualizar memoria
Français	synchroniser, rafraîchir mémoire
Deutsch	synchronisieren, Speicher aktualisieren

Usage

Automatic (after setup)

Just use Claude normally. Memory works in the background:

• Sessions are saved automatically every 3 exchanges
• New conversations load previous context on startup
• Memory refreshes every 5 exchanges to catch cross-session updates

Manual triggers in chat

• "sync" / "同步" / "メモリ同期" — Force sync with other sessions
• "save memory" / "保存对话" — Force save current session
• "what did Sonnet/Opus do?" / "Opus做了什么" — Check what another model discussed

File Structure

session-memory/
├── SKILL.md                  # Claude skill definition
├── README.md
└── scripts/
    ├── setup.sh              # One-time setup
    ├── save_session.py       # Core transcript parser & summary generator
    ├── cowork_save.py        # Runs inside Cowork VM, saves to Downloads
    ├── watch_save.py         # Host background daemon (10s poll, 30s save)
    └── sync_sessions.py      # Read and display saved memories

Requirements

• Python 3.8+
• No external dependencies (stdlib only)
• Works in Cowork, Claude Code, and Chat with Projects

License

MIT