what is hindsight-os fr?

Hindsight OS is a memory integrity layer built on top of Cognee, a graph and vector memory system. Where most memory tools for AI agents only add new information over time, Hindsight OS focuses on keeping that memory accurate: it remembers past decisions, recalls the relevant ones the moment new content is created, warns or blocks that content when it conflicts with what was already decided, learns from human feedback, and can deliberately forget information that is now obsolete, with a record of why. The project addresses a specific problem with AI memory. Decisions get superseded and incidents change the rules, but nobody re-reads the old record before repeating a past mistake. Worse, memory can be deliberately poisoned: the moment an agent writes to memory, a false fact can be planted, and once it lands there every future recall is compromised. The answer an organization already has, in an old decision record or incident report, often never resurfaces at the moment it is needed. Hindsight OS is seeded with a set of real engineering memories such as decision records, incident postmortems, and security standards, which Cognee turns into a graph based memory. For every new proposal, message, pull request, or agent action, the system recalls related evidence, traverses the knowledge graph for connected facts, and classifies the new content as a conflict, confirmation, duplicate, stale assumption, or something unrelated. A component called Sentinel acts as a firewall that detects attempts to poison memory, such as fake authority claims or fabricated approvals, and quarantines that content before it can be absorbed. Warnings or quarantine actions appear directly where the risk happens, for example a GitHub pull request comment or a live chat message, and the system learns from human feedback over time. The README describes this as one reusable pattern that could extend beyond engineering decisions to areas like newsroom fact checking, video platform moderation, community moderation, customer support, and compliance, though the README notes only the AI agent and engineering decision use cases are currently built, with the rest described as future vision.