dsh-memory

After a session ends, the agent forgets your project preferences, release constraints, and hard-won lessons. dsh-memory persists that long-term knowledge into plain Markdown files on your machine. On the next request the agent first sees a compact global memory plus a workspace index, and only opens the relevant detail file on demand through DSH file tools — no vector database, no dedicated memory service, and no flood of full history back into the context window.

The design blends two proven ideas: Claude Code-style progressive disclosure (index first, details as needed) and Codex-style session consolidation (extracting reusable signals from finished, stable sessions). Global memory holds cross-project preferences while workspace memory keeps project-specific facts isolated. A dedicated consolidation agent proposes memory changes, and deterministic host code handles ownership validation, concurrent-write protection, and the final commit. Everything is manageable from the Memory section in DSH Settings, with a bilingual interface that follows the system locale.

It is built for engineers who work daily in DeepSeek Harness and want memory that stays transparent and local-first: every Markdown file can be opened, edited, or backed up at any time; data lives on your machine by default; and there is no extra memory server to deploy. If you value privacy, inspectability, and lightweight cross-session context continuity, dsh-memory delivers a practical, zero-infrastructure solution.