dsh-hybrid-memory

Hooking long-term memory into DeepSeek Harness has always been a trade-off: Hermes-style snapshot files are token-frugal but unsearchable and unaudited, while Noema gives you full-text search and explainability at the cost of an external Rust service and prompt-mutating recalls that break the KV-cache prefix. dsh-hybrid-memory stitches both approaches into a single local pipeline so you can keep your prefix cache intact while still querying a rich, searchable fact store on demand.

The plugin works in three layers. L1 freezes MEMORY.md and USER.md into the system prompt at session start; any mid-conversation writes hit disk without touching already-assembled content, so the KV-cache prefix stays valid and the character budget (4000 for agent, 2000 for user) is model-agnostic. L2 builds a node:sqlite FTS5 index with a 2-character sliding-window pre-tokenizer for accurate Chinese search, and exposes memory_search, memory_browse, and memory_recall tools plus an optional review queue for accepting, editing, or rejecting candidate facts. L3 imports existing memories from Hermes, Claude Code, Codex, or WorkBuddy with a content-hash ledger that deduplicates automatically. All data stays in a local directory you choose, with no telemetry and no network calls.

It is well suited for developers running long, multi-session conversations in DSH who need persistent user-preference and project-fact recall, particularly in Chinese-language workflows where search accuracy matters, and for anyone migrating a memory library from another tool into a single, cache-friendly store.