What is RISI?
RISI is an agent-operable, human-governed security research harness for persistent AI-agent memory. It studies whether authorized retrieval changes adaptive state and whether ordinary, truthful interactions can make a legitimate critical memory unavailable at a later decision point. The standalone harness is primary. Humans and AI agents use the same versioned contracts, safety kernel, evidence pipeline, and verification commands. Ecosystem integrations such as MCP servers or skills may be added later as optional wrappers.Research status
RISI, RISI-C, CRAF, and memory eclipsing are proposed terms and hypotheses. They are not validated findings. The repository implements a deterministic safe baseline and one intentionally controlled synthetic CRAF mechanism with a protected control. It does not demonstrate an external vulnerability. The authoritative research charter is maintained separately in the private research vault. This documentation describes released repository behavior and public research boundaries only.Start here
- Quickstart — execute the DEP-01 baseline and controlled CRAF comparison.
- Operator guide — understand contracts, exit codes, and evidence.
- Core concepts — understand the terminology and boundaries.
- Responsible use — review authorization and safety requirements.
- Architecture — see the enforcement and evidence flow.