📦 Context Sandbox — 99.2% AI Context Compression

From 35,889 tokens to 298. Unlock unlimited conversation history.

Context Sandbox is a tool output compression engine that reduces agent context usage by 99.2%. Instead of loading raw tool outputs into every LLM call, Context Sandbox stores them in an FTS5-indexed sandbox and retrieves only what's needed — when it's needed.

• 99.2% Compression — Full tool outputs (35,889 tokens) → compressed summary (298 tokens)
• FTS5 Full-Text Search — SQLite-powered semantic retrieval across sessions
• "Think in Code" Paradigm — Let AI analyze programmatically instead of loading raw data
• Session Continuity — No more "I forgot what we did 3 turns ago"
• Tool Output Caching — Each unique output is stored once, referenced by hash

git clone https://github.com/Youan-ai/Context-Sandbox.git
cd context-sandbox

from sandbox import ContextSandbox

sandbox = ContextSandbox()

# Instead of loading 35K tokens into context:
summary = sandbox.compress(large_tool_output)
# → Returns 298 tokens with FTS5 index for retrieval

# Later, search across all compressed outputs
results = sandbox.search("What was that Python function?")
# Found instantly, no context waste

┌────────────────────────────────────────┐
│           Context Sandbox               │
├──────────────┬─────────────────────────┤
│ Compression   │ FTS5 Index              │
├──────────────┼─────────────────────────┤
│ hash dedup   │ SQLite full-text search  │
│ strip noise  │ cross-session retrieval  │
│ summarize    │ relevant snippets only   │
└──────────────┴─────────────────────────┘

Less context, more intelligence. Contributions welcome.

Apache 2.0 — see LICENSE.