dsh-native-memory

Juggling multiple projects and lengthy sessions inside DeepSeek Harness (dsh) tends to leave the model starting from scratch every turn or dumping the entire history into context — expensive and imprecise. dsh-native-memory solves this by persisting long-term facts and profiles directly in dsh s own storage domain, scoped per workspace (exact-cwd authorization), with no external server or extra runtime dependency beyond zod and the harness itself.\n\nIts design centres on safe, auditable memory. Every write (remember / edit / forget / import) passes through a human approval gate and lands in the session-log audit trail. Recall is tiered: an always-injected bounded profile, a deterministic three-tier keyword scan (tags, then text, then fuzzy match, with freshness and access-frequency tie-breaks), and full-text search over the current workspace s past sessions (the calling session excluded). Each fact carries its origin (sessionId, seq) so memory stays fully reconstructable from the lossless session log. Secret-shaped text (tokens, keys, passwords) is rejected by default, and prompt-injection payloads plus tool output are always masked. A read-only browser page (Settings > Memory) lists every workspace s facts with secrets masked, deletions are routed back through the approval gate, and an opt-in session-end proposal distills a finished session into candidate facts that only become real memories after manual approval. A compaction drift guard (on by default) deterministically surfaces literal anchors that a compaction summary dropped, as verification data in the next sessions — no LLM required.\n\nIt is aimed at developers maintaining multiple dsh workspaces who want per-project memory isolation, human-approved writes, full provenance, and zero external moving parts. If you are used to the global, silent write model of dsh-hermes-memory, think of dsh-native-memory as a more conservative, finer-grained alternative; if you are building a multi-project dsh workflow from scratch, it gives you a native memory foundation out of the box.