Deepseek-Harness-Lifelong-Agent

DeepSeek Harness agents accumulate large volumes of experience during inference, yet letting the model read and write that experience at will undermines both auditability and safety. DSH Governed Memory adds a governed memory layer on top of DSH: every record is persisted in an atomic-write JSON registry, and promoting a record to retrievable status requires explicit approval from a trusted host plugin backed by verified A/B evidence. The model itself cannot invoke capture or promotion tools, ruling out forged injection at the protocol level.

At retrieval time, the plugin screens each candidate against scope, privacy flags, expiration, conflicts, relevance, and a token budget; a single failed check denies the record. Revoked records are permanently excluded from all subsequent lookups. Every record must carry full structured metadata—evidence URIs, locators, proof level (A/B/C), and verification status—ensuring any retrieved memory remains traceable to its original source.

This is a technical preview limited to local and project scopes, running as a single process. It suits developers and security-governance teams who want an auditable, revocable, model-unforgeable experience memory inside DSH. Do not store credentials or raw private transcripts in the local registry file.