After reading about the Replit incident — where an agent deleted a production
database during a code freeze, then fabricated status reports — I kept thinking
about the root cause.
The agent didn't malfunction. It had no memory of what its past actions caused.
Current tools like Mem0 and Zep store what happened. None of them store what an
action caused downstream in a structured, queryable way. So the same mistake can
repeat indefinitely across sessions.
I built CausalOS to fix this. It's a causal memory layer that:
- Records action → outcome chains as a graph (not a vector blob)
- Runs semantic recall before any action: "has something similar caused damage before?"
- Blocks or warns via CausalGuard — fully deterministic, zero LLM calls
100% local. No API key needed. MIT licensed.
pip install causal-os
GitHub: https://github.com/CausalOS/causalos-python
Happy to answer questions about the architecture or the design decisions.
Still early — would genuinely appreciate feedback on the API surface.
Top comments (0)