dsh-plugin-xt-memory

DeepSeek Harness does not ship with cross-session memory by default, so every conversation begins from a blank slate. dsh-plugin-xt-memory is built for teams who have already accumulated project knowledge with XT-AGENT, or developers who want their assistant to automatically distill and reuse context across sessions. It ports the full core logic of XT-AGENT packages/memory into dsh's section-provider-plus-tool shape, meaning existing XT memory files work out of the box with zero migration.

On the injection side, a BM25 relevance score (with CJK 2-gram support), type weighting, and time decay select the top-K memories within a character budget and splice them into the system prompt, complete with an untrusted disclaimer line to keep injected content from being mistaken for instructions. The tool surface exposes memory_read, memory_search, and memory_write, with automatic sanitization and semantic de-duplication on every write. A background extraction pipeline intercepts each turn's primary messages, applies heuristic pre-filtering and a 60-second debounce, then calls a dedicated extraction model to persist key points, followed by semantic merge (write / merge / skip), all without interrupting the user.

Every memory carries hitCount and lastUsedAt telemetry; entries untouched for 30 days are auto-archived to an archive/ folder while the original text is preserved. Storage uses Markdown with frontmatter, fully compatible with the XT format, and environment variables such as XT_MEMORY_BUDGET_CHARS and XT_MEMORY_INJECT_TOP_K continue to work with dsh config taking precedence. This plugin fits developers who have already built a memory corpus in the XT-AGENT ecosystem, as well as teams looking to bootstrap cross-session self-learning for dsh from scratch.