DSH Noema

Long-term memory for DeepSeek Harness — durable, inspectable agent memory backed by Noema.
_{Recall Before Work • Import From 9 Agent Tools • Settings-Page Memory Management • Crash Keep-Alive • Hot Reload}

_{npm: @zseven-w/dsh-noema · Current plugin release: 0.1.0-rc.1 · Tested with DSH 0.1.0-rc.6}

English · 简体中文 · 繁體中文 · 日本語 · 한국어 · Français · Español · Deutsch · Português · Русский · हिन्दी · Türkçe · ไทย · Tiếng Việt · Bahasa Indonesia

DSH Noema — memory settings page

_{The Noema Memory settings page — import sources, memory management, and live server status}

Why DSH Noema

DSH Noema connects DeepSeek Harness with Noema — a local-first, non-vector memory system for coding agents — so an Agent keeps durable knowledge across sessions instead of starting every conversation from zero.

🧠 Durable Recall

Memories persist as inspectable Markdown files under NOEMA_ROOT (default ~/.agent-memory/). noema_recall loads relevant context at the start of a session; noema_search, noema_browse, noema_catalog, and noema_recall_graph cover lookup, exploration, and auditing.

📥 Import From Other Tools

noema_import reads the memory files of nine other AI coding tools — Codex, Claude Code, opencode, Cursor, Grok, WorkBuddy, Antigravity, Trae, Qoder — splits them into sections, and saves each as a durable memory. A content-keyed ledger deduplicates across runs and across tools that share files.

🛠️ Settings-Page Management

The Noema Memory settings page configures the server command, memory root, budgets, idle/call timeouts, and the guidance section — and a Manage memories card searches, browses, adds, reviews, and deletes stored memories directly.

🩺 Keep-Alive

The memory server stays up: idle timeout defaults to never, and a keep-alive loop restarts the noema-mcp child in the background when it crashes or exits, with a configurable check interval and restart backoff.

🔍 Smart Entity Extraction

Noema's extraction engine combines jieba word segmentation with high-precision signals — English proper nouns, CJK names and technical terms, quoted topics, and repetition — with stopword and path filters, so the PageIndex topic catalog stays clean.

⚡ Hot Reload

After the first boot, the plugin never needs a restart again: pnpm run build hot-reloads the host plugin through Cordis HMR, and ppnpm run build:client hot-swaps the browser bundle over the client-hmr SSE channel.

Install into DSH

dsh plugin --profile web add @zseven-w/dsh-noema@latest
dsh web

Or, for local development straight from the source tree:

dsh plugin --profile web add link:/path/to/dsh-noema
dsh web

The link: protocol symlinks the profile dependency to this repository, so rebuilds are visible immediately and Cordis HMR can watch the compiled output.

The plugin bundles the noema-mcp binary through per-platform optional npm packages. To build it yourself instead, run cargo build --release -p noema-mcp inside the bundled noema submodule, or point the Server command setting at any noema-mcp build.

Memory Tools

The model-facing tools mirror the Noema MCP surface:

Tool	What it does
`noema_recall`	Recall relevant memories for a query, with a token budget.
`noema_search`	Full-text search over stored memories.
`noema_browse`	Browse the PageIndex catalog for a topic or entity.
`noema_catalog`	Render the full memory catalog as markdown.
`noema_recall_graph`	Multi-hop recall through links and shared entities.
`noema_neighbors`	One graph hop from a memory.
`noema_explain`	Explain why a memory was or was not recalled.
`noema_remember`	Save a durable fact, decision, constraint, or preference.
`noema_review_list`	List pending review candidates.
`noema_review_decide`	Accept, reject, edit, or merge a candidate.
`noema_forget`	Tombstone or hard-delete a memory.
`noema_policy_get` / `noema_policy_set`	Read or update the write policy.
`noema_status`	Server and tenant status: counts, index health, storage root.
`noema_import`	Import memories from other AI coding tools.

Each tool returns a uniform envelope { ok, tool, text } where text carries the full server output.

Import memories from other tools

Source id	Global files	Workspace files
`codex`	`~/.codex/AGENTS.md` + the Codex memory pipeline: `~/.codex/memories/MEMORY.md`, `memory_summary.md`, `rollout_summaries/.md`, `extensions/ad_hoc/notes/.md` (`raw_memories.md` skipped — it is the uncurated feed)	`AGENTS.md`, `AGENTS.local.md`
`claude-code`	`~/.claude/CLAUDE.md`, `~/.claude/CLAUDE.local.md`, `~/.claude/MEMORY.md`	`CLAUDE.md`, `CLAUDE.local.md`, `MEMORY.md`
`opencode`	`~/.config/opencode/AGENTS.md`	`AGENTS.md`
`cursor`	`~/.cursor/rules/*.mdc`, `~/.cursorrules`	`.cursor/rules/*.mdc`, `.cursorrules`
`grok`	`~/.grok/AGENTS.md` + the Grok cross-session memory: `~/.grok/memory/MEMORY.md`, per-project `MEMORY.md`, and `sessions/*.md` summaries	`AGENTS.md`
`workbuddy`	`~/.codebuddy/CODEBUDDY.md` (WorkBuddy memory file), `~/.workbuddy/AGENTS.md`, `~/.workbuddy/memory.md`, `~/.config/workbuddy/AGENTS.md`, `~/Library/Application Support/WorkBuddy/AGENTS.md`	`AGENTS.md`, `CODEBUDDY.md`
`antigravity`	`~/.antigravity/AGENTS.md`, `~/.config/antigravity/AGENTS.md`, `~/Library/Application Support/Antigravity/AGENTS.md` (best-effort; no documented global memory store yet)	`AGENTS.md`, `AGENTS.local.md`
`trae`	`~/.trae/AGENTS.md`, `~/.trae/memory/`, `~/.trae/rules/` (plus the `~/.trae-cn` variants)	`AGENTS.md`, `.trae/rules/`
`qoder`	`~/.qoder-cn/AGENTS.md`, `~/.qoder-cn/rules/`, the auto-memory roots `~/.qoder-cn/memory/` and `~/.qoder-cn/projects/*/memory/` (plus `~/.qoder` variants)	`AGENTS.md`, `AGENTS.local.md`, `.qoder/rules/`

The source argument selects one tool, or omit it to run every source enabled in settings.
The path argument selects the workspace root for project-scoped files (defaults to the session workspace; workspace files only load when the Import workspace files setting is on).
Imports are deduplicated through a ledger at $DSH_HOME/storages/dsh-noema-imports.json, keyed by file path + section content — when several tools share one project AGENTS.md, each section is imported exactly once. force: true re-imports everything.
The settings page exposes per-source checkboxes, an import-on-startup toggle, a file-size cap, and an Import now button with a last-run summary.

Settings

Open Settings → Noema Memory:

Setting	Default	Meaning
Enable memory	on	Master switch for the `noema_*` tools.
Memory guidance	on	System-prompt section teaching memory usage.
Start server at boot	on	Spawn at DSH start instead of first use.
Auto-accept new memories	on	`noema_remember` persists immediately.
Server command	`bundled`	Bundled `noema-mcp` binary or a custom executable path/command.
Working directory	—	cwd for the server (needed for `cargo run`).
Memory root (NOEMA_ROOT)	—	Where memories are stored; empty = `~/.agent-memory`.
Recall token budget	1200	Default `budget_tokens` for `noema_recall`.
Idle timeout (ms)	0	Stop the server after idle; 0 = never.
Keep alive	on	Restart the server in the background when it crashes or exits.
Keep-alive interval (ms)	5000	Minimum delay between background health checks.
Call timeout (ms)	30000	Per-tool-call deadline.
Restart delay (ms)	1000	Backoff between a stop/crash and the next start.

The status card shows server health with restart/stop actions, and the import section manages the nine memory sources.

Hot reload

DSH's HMR machinery is fully usable once the plugin has been loaded once:

Host plugin — enable the Cordis HMR entry in the profile patch with its watch root pointed at this package's lib/ output, and keep the link: dependency. Run pnpm run build and the running DSH reloads the plugin entry automatically (the Noema server child is restarted by the reload) — no server restart.
```
# ~/.dsh/profiles/<profile>/cordis.patch.yml
- id: hmr
  disabled: false
  config:
    root:
      - /path/to/dsh-noema/lib
```
Client bundle — ppnpm run build:client rewrites lib/client.js; the client-hmr node half stat-polls every graph bundle (default 500ms) and broadcasts a rebuilt frame over the /plugins/events SSE channel, and the browser hot-swaps the module without a page refresh.
Settings — every change made on the Noema Memory settings page applies live through the settings service.

The one thing hot-reload cannot do is load a plugin that was never in the booted tree: the running composition neither watches the profile patch layer (the web app does not wire watchUserPatches) nor exposes a loader mutation API (the plugin inventory RPC is read-only). A fresh plugin therefore needs exactly one server restart, after which the loop above is fully hot.

Develop

pnpm install
pnpm run build     # host tsc + client tsdown bundle
pnpm test          # build + node --test tests/

The e2e test runs against noema/target/debug/noema-mcp when present (it is skipped otherwise).

License

MIT