dsh-native-memory

Native, per-workspace long-term memory for DeepSeek Harness (dsh): facts and profiles stored on the harness's own storage-domain, cross-session recall through session-query FTS, approval-gated writes, and cited provenance — no external server, no extra runtime dependency.

⚠️ Installing a plugin runs third-party code on your machine with your own permissions. Review the source (src/) and the security model before installing.

Why this one

	dsh-hermes-memory	dsh-native-memory
Storage	~/.dsh/settings.yaml namespace	dedicated storage-domain unit (~/.dsh/storages/dsh_memory.json)
Scope	user-global, all projects	per workspace (exact-cwd authorization)
Write safety	silent, model-only	human approval gate + session-log audit
Recall	everything always injected (hard caps)	bounded always-on profile plus on-demand recall + FTS over past sessions
Dependencies	vendored imports into the harness checkout	none beyond zod + the harness itself

See docs/design.md for the full architecture and the competitive landscape analysis.

Install

dsh plugin --profile web add dsh-native-memory   # npm after release
dsh plugin --profile web add /path/to/this/repo  # from a checkout

Restart dsh web. The bundle enables session-query full-text search and adds the memory tools to every session. Details and configuration: docs/install.md.

Tools

Tool	Kind	Gate
memory_remember	add/update a fact in this workspace (secrets rejected by default)	approval
memory_edit	replace a fact	approval
memory_forget	archive a fact (soft delete)	approval
memory_recall	deterministic three-tier keyword scan (tags > text > fuzzy; freshness/access tie-breaks)	none
memory_search	FTS over this workspace's past sessions (caller excluded)	none
memory_expand	expand a fact's citation to the original log excerpt	none
memory_consolidate	near-duplicate merge suggestions + cap budget	none
memory_import	import candidate facts from a past session's log	approval (per fact)
memory_profile	read the always-injected workspace profile	none
memory_export	write a git-friendly Markdown mirror (.dsh-memory/memory.md, masked, idempotent)	none

Every fact records its origin (sessionId, seq) — memory stays reconstructable from the lossless session log.

Writes reject secret-shaped text (tokens / keys / passwords) by default; secretPolicy: "mask" | "off" in the bundle patch relaxes that. The credential-assignment detector can flag benign token: … values of ≥16 characters. Prompt injection and tool output always mask secrets.

A read-only browser page (settings → 记忆) lists every workspace's facts with secrets masked; deletions are copied as a memory_forget instruction and land in the chat through the approval gate.

Opt-in session-end proposals (proposeOnSessionEnd: true): one cheap LLM call distills a finished session into candidate facts shown in the next sessions; they become facts only through the approval-gated memory_remember.

A compaction drift guard (compactionGuard: true, on by default) surfaces literal anchors a compaction summary dropped, as data to verify in the next sessions — deterministic, no LLM.

Development

pnpm install
pnpm build && pnpm typecheck && pnpm test

New contributors start at docs/handoff.md and docs/contributing.md. Chinese docs: README.zh.md.

License

MIT