aimem-cli 0.2.0

AI Memory Switcher — Save, compress, and transfer context between AI agents

🧠 AiMem — AI Memory Switcher

Save, compress, and transfer context seamlessly between AI agents. When Claude hits rate-limits or you're mid-flow, AiMem bridges the gap — without losing your train of thought.

Agent A (Claude)  ──▶  Read Session  ──▶  Universal Format  ──▶  Save  ──▶  Agent B (Gemini)
  (JSONL)              (.jsonl)             (JSON)              (~/.aimem/)     (Markdown/Prompt)

Zero server. Zero Redis. Works offline. LLM compression is opt-in.

---

Install

pip install aimem-cli

Or from source:

git clone https://github.com/ThangTo/AiMem.git
cd AiMem
pip install -e .

---

Quick Start

# 1. Save current Claude session
aimem save --from claude

# 2. Load into Gemini (auto-copied to clipboard)
aimem load sess-abc123 --to gemini

# That's it. Paste into Gemini CLI.

---

Commands

aimem init                          Initialize config (~/.aimem/config.json)
aimem save --from claude            Save session (interactive if multiple)
aimem save --from clipboard         Save from system clipboard
aimem save --from qwen              Save from Qwen CLI
aimem save --from gemini            Save from Gemini CLI
aimem save --from aider             Save from Aider chat history
aimem save --from continue          Save from Continue.dev
aimem load <id> --to gemini          Load session as target format
aimem list                           List saved sessions
aimem list --agents                  Check which agents are available
aimem config                         Show current config
aimem config set key=value           Update config
aimem delete <id>                    Delete a saved session

Source agents (--from)

Agent	Storage	Inject?
claude	~/.claude/projects/*/*.jsonl	✅
gemini	~/.gemini/tmp/*/chats/*.json	✅
qwen	~/.qwen/tmp/*/logs.json	✅
opencode	~/.opencode/sessions/*.json	✅
codex	~/.codex/sessions/*.jsonl	✅
aider	~/.aider.chat.history.md	❌
continue	~/.continue/sessions.db	❌
clipboard	System clipboard	❌

Target formats (--to)

Format	Best for	Inject?
markdown	Paste into any web UI or tool	❌
claude	Claude Code CLI	✅
gemini	Gemini CLI	✅
qwen	Qwen CLI	✅
opencode	OpenCode CLI	✅
codex	Codex CLI	✅
continue	Continue.dev (VS Code)	❌
prompt	API calls / custom injection	❌

Auto-Inject (NEW!)

Inject directly into target agent storage — no copy-paste needed:

# Inject into Gemini (appears in session list)
aimem load claude-62d520bb --to gemini --inject
gemini --resume latest

# Inject into OpenCode
aimem load claude-62d520bb --to opencode --inject
opencode -s ses_xxx

# Inject into Claude
aimem load claude-62d520bb --to claude --inject
claude --resume

---

Configuration

Config: ~/.aimem/config.json

# Show config
aimem config

# Enable LLM compression (opt-in)
aimem config set compression.enabled true
aimem config set compression.api_key YOUR_GROQ_KEY

# Switch compression provider
aimem config set compression.provider groq
aimem config set compression.provider gemini

# Change output format
aimem config set output.format markdown
aimem config set output.clipboard_auto true

# Enable Redis cache (optional)
aimem config set storage.redis.enabled true
aimem config set storage.redis.host localhost
aimem config set storage.redis.ttl 3600

---

Architecture

aimem/
├── aimem/
│   ├── __init__.py
│   ├── cli.py              # CLI interface + all commands
│   ├── models.py           # UniversalSession, Message, CompressedSession
│   ├── storage.py          # FileStorage (default) + RedisCache (opt-in)
│   ├── compression.py      # LLM Compression Engine (Groq / Gemini)
│   └── adapters/
│       ├── claude.py       # Read ~/.claude/projects/*/*.jsonl
│       ├── qwen.py          # Read ~/.qwen/tmp/*/logs.json (gzip)
│       ├── gemini.py        # Read ~/.config/gemini/ (JSON/JSONL)
│       ├── aider.py         # Read ~/.aider.chat.history.md
│       ├── continue_dev.py  # Read ~/.continue/sessions.db (SQLite)
│       ├── clipboard.py     # Read system clipboard
│       └── output/
│           └── __init__.py  # Markdown, Claude, Gemini, Qwen, Continue, Prompt
└── aimem_main.py           # Entry point (run without install)

Universal Session Format

Every agent's session is converted to this neutral JSON format:

{
  "id": "claude-62d520bb",
  "source": "claude",
  "messages": [
    {"id": "...", "role": "user", "content": "...", "timestamp": "..."},
    {"id": "...", "role": "assistant", "content": "...", "timestamp": "..."}
  ],
  "context_items": [],
  "compressed": {
    "current_goal": "Build context transfer tool",
    "latest_code": [{"path": "cli.py", "content": "...", "language": "python"}],
    "current_errors": ["Error: undefined is not a function"],
    "key_decisions": ["Use file-based storage over Redis"],
    "todo_list": ["Write Continue.dev adapter", "Add compression"]
  },
  "metadata": {
    "source_agent": "claude",
    "original_session_id": "62d520bb",
    "project_path": "D:\\Project\\AiMem",
    "model": "claude-sonnet-4-6"
  },
  "created_at": "2026-04-18T08:11:00Z"
}

LLM Compression (Opt-in)

When --compress is used (or compression.enabled is true in config):

Input: ~114k tokens (Claude session) + Groq API call
                          ↓
Output: ~2k tokens (compressed summary)
                          ↓
Save to ~/.aimem/sessions/ as JSON

Providers:

• Groq llama-3.1-8b-instant — Fast + Cheap (~$0.001/session)
• Gemini gemini-2.0-flash-exp — Google's fast model

Requires compression.api_key to be set.

---

Workflow Example

# You're coding in Claude, session hits limit
$ aimem save --from claude
[i] Found 78 Claude sessions. Select one:
  [1] 2026-04-18T08:11 | Tôi muốn tạo một tool giúp chuyển context...
  [2] 2026-04-17T14:46 | Research how different AI CLI agents...
Enter number (default=1): 1
[OK] Saved session: claude-62d520bb

# Switch to Gemini, paste context
$ aimem load claude-62d520bb --to gemini
## Previous Context

Continuing from previous session:

Project: `D:\Project\AiMem`
**Goal:** Build AiMem - context transfer tool

**Key Decisions:**
- File-based storage (no Redis dependency)
- Python-first for AI/ML ecosystem
- Opt-in compression (not required)

**Todo:**
- [ ] Write Continue.dev adapter
- [ ] Add session interactive selection

---
**Continue from here.**

# Copied to clipboard automatically

---

Roadmap

Phase	Status	Description
Phase 1	✅ DONE	MVP — Claude, Qwen, Gemini, Clipboard, Aider, Continue adapters
Phase 2	⚠️ PARTIAL	LLM Compression — Groq works, Gemini blocked (API issue)
Phase 3	✅ DONE	Output adapters for all agents
Phase 4	🔜 NEXT	Smart chunking + context window management
Phase 5	📋 PLANNED	VS Code Extension
Phase 6	📋 PLANNED	GUI (optional TUI mode)

---

Design Decisions

Why File-Based by Default?

Approach	Setup Time	Portability	User Friction
Redis	5-10 min	Low (server-dependent)	High
File (AiMem)	0 min (works immediately)	High (copy config file)	Low

AiMem saves sessions as plain JSON in ~/.aimem/sessions/ — no server, no daemon, no Redis. Transfer a session from your laptop to a server with one aimem load + aimem save.

Why Python?

• Python is pre-installed on most developer machines
• Native ecosystem for AI/ML tools (LLM APIs, SQLite parsing, etc.)
• Easy to extend with pip install aimem
• Native packaging via pyproject.toml

Why LLM Compression is Opt-in?

• Not every user has an API key
• Raw transfer works fine for most cases
• Compression adds latency (1-3s per save)
• Non-deterministic — may lose nuance

---

License

MIT