EternalNight996/memory-eternal

Auto-captures knowledge after every conversation and survives across sessions; recall fetches only the relevant chunks — saves tokens, less noise. Fully self-built, zero third-party memory framework, no DSH source changes, one vault shared by every Agent, SQLite persistent storage, zero external dependencies.

---

🚀 Get Started in 5 Minutes

🟦 DeepSeek Harness (DSH) — focus

Install (official desktop app via the plugin marketplace, dsh web via one CLI line):

# 1) Official DeepSeek Harness desktop app (recommended): Settings -> Plugins -> search memory-eternal
#    Equivalent CLI (that shell's profile is named "desktop"):
dsh plugin --profile desktop add memory-eternal

# 2) dsh web (browser):
dsh plugin --profile web add memory-eternal

# or update directly in the profile (pnpm workspace). Use pnpm — `npm install` throws EUNSUPPORTEDPROTOCOL
cd ~/.dsh/profiles/web && pnpm add memory-eternal@latest

# right after a release, @latest can be held back by pnpm's minimum-release-age gate or a cached
# packument (it silently stays on the old version): pin the version instead — pnpm adds it to the
# workspace's minimumReleaseAgeExclude automatically
cd ~/.dsh/profiles/web && pnpm add memory-eternal@0.9.6

# verify the version, then restart the host
node -e "console.log(require('memory-eternal/package.json').version)"

🖥️ Desktop support policy (read this first)

• The official DeepSeek Harness desktop build is the priority — and the only guaranteed target (the official Electron release; its profile is named desktop). The v0.9.4 → v0.9.6 fixes all target it: the 0.1.7 settings-service rewrite, schemastery 3.18.4 volatile live references, and window-titlebar avoidance (data-windows-titlebar, so a full-screen overlay can no longer cover minimize/maximize/close).
• The author's own dsh-desktop (dsh-desktop-shell / dsh-ui-* / dsh-plugin-marketplace) is discontinued and is no longer a support target — the official desktop build covers the same ground, and maintaining two shells only splits compatibility. Note that its profile still pins dsh-desktop-shell@0.3.0, which has been unpublished from npm and makes pnpm install fail outright; migrating to the official desktop build and dropping those dependencies is recommended.
• Compatibility matrix (both verified): the official desktop profile (schemastery 3.18.4) and the web profile behind dsh web (schemastery 3.18.1) share the same code path.

After restarting the host, three things are live immediately:

Effect	Where
Auto-capture of knowledge cards	Happens every turn, no action needed
memory_recall tool	Agent calls it automatically when it needs history
Visual UI	Sidebar bottom Memory button / Settings → Memory

Entry: The sidebar-footer Memory button opens the vault (its left rail includes Cards / Graph / Usage / Audit Center / Recycle Bin / Memory Config); "DSH Settings → Memory" is the pure config page.

UI language: The whole UI is bilingual (English / 中文) and follows DSH Settings → Language in real time — Cards, Graph, Audit Center, card templates, config panel all included; the standalone web page follows the browser language.

Edit config: Vault left rail "Memory Config" (or DSH Settings → Memory) → DSH memory config / cost control / auto-audit config / self-hosting — just hit "Save Config". autoWebMode / watchdogAutoSpawn changes require a DSH restart.

🟨 Claude Code

npm i -g memory-eternal     # installs CLI + MCP (auto-writes ~/.claude.json + SessionEnd hook)

Ready after install: say "recall 数据库选型" in a session → auto-retrieves memory; on session end / before context compaction → auto-captures.

🟧 Codex CLI / Cursor

npm i -g memory-eternal     # auto-writes Codex config.toml / Cursor mcp.json

Restart the tool → MCP is in the list; just use: 用 memory_recall 查一下项目历史决策.

🟩 Browser (no Agent needed)

dsh-memory open    # starts web + opens browser (default http://127.0.0.1:7999)

Stats / search / card grid (add-edit-merge-import-export) / knowledge graph — all here. Same UI as the DSH embed; data stays in sync.

zcode (Zhipu): no native MCP; bridge via zcode-open-bridge or use the CLI directly.

---

📖 Command Reference

dsh-memory recall "database selection"   # retrieve
dsh-memory capture "important note..."   # manual capture (- reads stdin)
dsh-memory sweep ~/.claude/projects      # mine existing sessions
dsh-memory setup [--dry-run]             # re-run / preview auto-mount (idempotent)
dsh-memory mcp                           # MCP stdio (mount to any MCP client)
dsh-memory serve [--port 7999]           # run web in foreground
dsh-memory open                          # ensure web alive + open browser
dsh-memory watchdog [--port 7799]        # watchdog keep-alive (standalone process)

When not running on DSH, the dsh-memory command comes from npm i -g.

---

⚙️ Self-hosting (plain words)

Three concepts, don't mix them:

• How the web server stays alive (autoWebMode) → init=pull once at DSH start (default); interval=DSH in-process timer probes & auto-restarts (0 extra memory); manual=fully manual, only from dsh-memory open.
• Watchdog process (watchdogAutoSpawn, default on) → a standalone node process that can pull web up even after DSH exits (~+47 MB RAM). Only turn on for 7×24 keep-alive.
• Auto-mount MCP (autoMcpSetup, default off) → whether to auto-write MCP into Claude Code/Codex/Cursor config. Off = don't touch your machine config; run dsh-memory setup manually when needed.

Change these: Vault left rail "Memory Config" (or DSH Settings → Memory) → edit the table and save; autoWebMode / watchdogAutoSpawn need a DSH restart.

MCP is a protocol, not a resident service: the agent spawns it per session and exits when done — no "auto-start on boot" concept.

Three deployment intensities

Scenario	Config	Memory
Personal dev (default)	autoWebMode=init + watchdogAutoSpawn=off	web 47 MB
Resident 7×24	watchdogAutoSpawn=on	web + watchdog 47+47 MB
True boot auto-start (no DSH)	Windows Task Scheduler runs dsh-memory watchdog --port 7799 --interval 5000 --max-restart 10	same

---

⚙️ Memory Config (plain words)

All settings live in the vault left rail "Memory Config" (or DSH Settings → Memory) — edit and hit "Save". Here's the full page (plugin info / Agent MCP mount status / auto-audit config):

1. Most used

Setting	Default	In plain words
Auto capture	on	auto-store useful content as cards after each turn
Auto recall	on	AI auto-queries memory when it needs history
Vault dir	~/.dsh/memory-vault	where memory lives, plain Markdown & git-able

Multiple vaults (per-project isolation)

Two extra fields in memory-eternal-config.json:

{
  "vaultProfiles": [{ "name": "work", "path": "D:/vaults/work" }],
  "activeVault": "work"
}

Resolution order: MEMORY_VAULT_DIR env → the path of the vaultProfiles entry whose name matches activeVault → vaultDir → default ~/.dsh/memory-vault.

Since v0.9.8 the CLI / MCP / hooks / standalone web / sweep share the host's exact order — previously only the host honoured activeVault, so switching vaults left the terminal and hooks writing to the default vault and split your memory in two.

Note: the settings page still has no form for vaultProfiles / activeVault (edit the config file by hand), and per-workspace auto-selection is not implemented yet.

2. Save money (important)

Setting	Default	In plain words
Distill cards	on	compress a conversation into a sharp card (calls AI, costs money). Off = store raw text, zero cost
Dedup feeds AI	on	judge if new content is a duplicate (calls AI). Off = simple dedup, saves one AI call
Distill output cap	900	max chars per compress, bigger = sharper but pricier
Recall min score	2	how "close" a match must be to return; bigger = sharper but leaks more (cheaper)
Min capture chars	200	too-short chats aren't stored (avoids small-talk waste)
Daily quota	60	max cards per day, prevents AI burning money

3. How the service runs

Setting	Default	In plain words
Keep-alive autoWebMode	init	init=open web once at DSH start; interval=periodically check & restart if dead; manual=fully manual
Watchdog watchdogAutoSpawn	on	a standalone process keeps web alive (+47 MB). Personal use can turn off
Auto-mount MCP autoMcpSetup	off	lets Claude Code / Codex / Cursor use your memory. On = auto-configures them; Off = never touches your config, run dsh-memory setup manually

💰 To save money: turn Distill cards off, lower Distill output cap, raise Recall min score.

🎯 One-click presets

Top of the config page: 🟢 A Light / 💰 B Budget / ⭐ C Premium — click to fill, then Save:

Plan	Scenario	Keep-alive	Watchdog	Distill	Distill cap	Recall min	Memory	LLM cost
🟢 A Light	Personal dev (default)	init	off	on	900	2	~47 MB	normal
💰 B Budget	Tight budget / many Agents	init	off	off	500	3	~47 MB	~0
⭐ C Premium	Long projects / teams	interval	on	on	1200	1	~94 MB	high

---

🔔 Capture Log & Failure Alerts

Auto-capture is a silent background pipeline: when it broke, the only symptom used to be "no new cards" with no visible cause. Since v0.9.0 it speaks up.

• Capture log: Settings → Memory → Usage / Today shows the last 100 run entries: 🚀 boot / 👂 session attached (which event API + event count) / ✅ card created / ➕ update appended / ⏭ skipped (reason) / ❌ failed (reason).
• Active alerts: any failure turns the strip red at the top of the memory library and injects "auto-capture failed + reason" into the agent context, so the agent tells you in its first sentence — no need to dig through logs.
• Stall watchdog: if turns start but no turn-stopping event arrives for 15 minutes (e.g. DSH renamed the event), a 10-minute sweep reports it.
• Self-adapting event API: the session event API is detected automatically (ownEvents() → snapshotEvents() → events), so a DSH upgrade that changes the API can no longer silently break capture.

---

🛡️ Audit Center & Recycle Bin

Agents cannot approve on your behalf (since v0.9.1). Whenever audit is active, the plugin injects a red line into the agent context: new cards stay pending, and calling the approve/reject endpoints (or any equivalent bypass) is forbidden. The decision is always yours.

New cards go to the Audit Center (pending) by default and enter the main vault only after you approve them; rejected ones move to "Rejected", where you can restore or delete to the recycle bin. Cards matching an exemption (audit mode = skip all / exempt agents / exempt kinds) go straight in; recycle-bin cards are recoverable within 30 days, then auto-purged.

• Pending / Rejected tabs, filtered by type / date / agent; select all and approve / reject / delete-to-recycle in one click.
• Rules in "Memory Config → Auto-audit config": audit mode (audit all / skip all) + exempt agents + exempt kinds + recycle retention days.

🥇 Why the audit system is more reliable (vs other memory products)

Most memory products (mem0 / Zep / agentmemory…) auto-ingest everything the moment a conversation ends, good or bad — noise, wrong facts, and sensitive content all go in and get recalled later, polluting context and amplifying hallucinations.

memory-eternal uses human-in-the-loop auditing — only trusted content enters the main vault:

Dimension	Other memory products	memory-eternal audit system
Ingest	Auto-ingest all, no gate	New cards first enter the Audit Center (pending), approved before entering the vault
Quality	No filtering of noise/errors	Only cards you've confirmed → more accurate recall, less noise
Traceability	No source/audit trail	Every card carries submittedBy author + pending/approved/rejected state
No-friction	—	Exempt agents / exempt kinds hit → trusted cards go straight in, zero wait
Safety	Delete = gone forever	Recycle-bin soft delete, recoverable within 30 days

---

🧬 Why Fully Self-built

Most memory solutions lean on third-party frameworks / spin up an MCP service / lock memory in a private store. This plugin builds the skeleton itself — zero third-party runtime deps, logic readable line by line:

Module	Self-built	Replaces
Dedup	lexical Jaccard bigram (0.62) + semantic dedup	duplicate-card prevention
Retrieval	CJK-aware: Chinese whole-word + char bigram	no full-text search engine needed
Graph	force-directed + [[wikilink]]/shared-tag edges	see knowledge links at a glance
Storage	plain .md with frontmatter	not locked in, readable, git-able, any tool can read

Same direction as popular projects (summarize→store→recall), but positioned differently: local, self-built, zero-dep, readable & controllable. If you already use mem0/Zep etc., just layer this as a "local persistent memory base".

---

🛠 Development / Test

npm i
npm test        # unit tests: vault dedup/retrieval/graph + capture pipeline + API shapes
npm run build   # builds lib/client.js (DSH embed) + web/app.js (standalone web bundle)

---

🐞 Reporting a bug

The left rail has a 🐞 Report a bug entry: describe the problem, then either

• Open prefilled issue on GitHub — title, body, diagnostics and the bug label are filled in; one more click on Submit files it;
• Copy AI prompt — paste it into your AI chat and it will run gh to de-duplicate, file the issue and report the URL back.

Why not one-click submission? GitHub forbids anonymous issue creation, a token inside the plugin would be extracted by decompiling, and OAuth / a relay needs hosted infrastructure — hence "pre-fill + one click". Diagnostics are auto-redacted (home directory → ~; sk- / ghp_ / github_pat_ / api_key: → ***), but do skim them before posting publicly.

If you cannot reach the UI, paste the prompt below into your AI chat and replace the last placeholder with your problem:

I hit a problem while using memory-eternal (a DeepSeek Harness memory plugin). Please file it on GitHub for me.

Repo: https://github.com/EternalNight996/memory-eternal

Workflow:

1. Run `gh issue list --repo EternalNight996/memory-eternal --state open --limit 50` first: if a matching issue exists, add a comment with `gh issue comment <n> --body-file <file>` instead of opening a duplicate.
2. Otherwise create it with `gh issue create --repo EternalNight996/memory-eternal --title "[Bug] <one-line summary>" --body-file <file>`, using sections: Symptoms / Steps to reproduce / Expected / Actual / Environment (plugin, DSH, OS, Node) / Diagnostics.
3. If `gh` is not authenticated or unavailable, do NOT try to submit: output a GitHub new-issue link with the title and body pre-filled instead, and I will click submit myself.
4. Report the resulting issue URL back to me.

Keep my description verbatim - do not rewrite or summarise it. Put the diagnostics under the Diagnostics section as-is.

My description:
<write your problem here>

Diagnostics:
<paste the diagnostics copied from the in-app Report-a-bug dialog>

---

📋 Changelog

Withdrawn versions: v0.9.16 – v0.9.22 are all withdrawn (problem-period releases: git tags deleted, npm entries deprecated — use v0.9.23). v0.9.14 is deprecated (its client-rendering changes hurt the feel and were reverted; its server-side algorithm work shipped in v0.9.15, whose rendering is identical to v0.9.13). v0.9.0 – v0.9.5 are all deprecated on npm, and the release tags for v0.9.1 / v0.9.2 / v0.9.3 / v0.9.4 have been removed from git (v0.9.5 is merely superseded, so its tag stays) — v0.9.1/v0.9.2 ship the blank-memory-page defect, v0.9.0 false-alarms its capture warnings, v0.9.3 does not support DSH v0.1.7-rc.2 (the plugin stops mounting entirely: the settings service changed and the client-side settingsScope is gone), v0.9.4 crashes under the official desktop build (schemastery 3.18.4 volatile live references) with cfg.vaultDir.trim is not a function, and v0.9.5's full-screen overlay covers the desktop shell's window controls (its top-right "×" sits on the window Close button — one click quits the whole shell). Use v0.9.6+ (npm i memory-eternal@latest). DSH compatibility: >=0.1.5-alpha.2 <0.2.0. v0.9.4 adds support for DSH v0.1.7-rc.2, which replaced the settings service with a forms-only API and removed @deepseek-ai/dsh-client-runtime / settingsScope; v0.9.5 adds support for schemastery ≥3.18.4 volatile live references (the shape the official desktop profile uses, while the web profile still ships 3.18.1 — both work); v0.9.6 keeps full-screen overlays clear of the desktop shell's window controls. The 0.1.5 line still uses the legacy settings.register path; both paths are kept.

Version	Date	Highlights
v0.9.23	2026-09-30	No more save flicker. After a successful save the client immediately re-fetches /config, but the host volatile echo lags, so inputs visibly reverted to the old values and then jumped back. The client now keeps a local overlay for fields that were saved but not yet echoed (src/client/config-merge.js, pure function + 4 checks), dropping each key once the host echoes it, and following the host value afterwards; resetting the form clears the overlay. npm test now runs 139 checks

Version	Date	Highlights
v0.9.22	2026-09-30	Near-instant config sync (SSE push + file watching). ① A new SSE hub (lib/sse.js): both hosts (DSH-embedded and standalone web) expose /memory-eternal/api/events, so a config change is pushed to every open page and the client reloads automatically — without overwriting what you are currently typing (dirty flag). ② Standalone save → DSH apply now triggers via fs.watch in milliseconds instead of a 5-second poll (the poll remains as a fallback). ③ The standalone page watches the shared config file, so a DSH-side save shows up there immediately. Adds tests/sse.test.mjs (5 checks); npm test now runs 135 checks
v0.9.21	2026-09-30	A successful save is no longer shown as a failure + config layout fix. ① Success always uses the success style (the "written, host has not echoed yet" case used to pop up in the red failure style, which read as "save failed"), reworded to "saved to the config file; restart DSH to make it effective"; ② config field grid widened from 150/160px to 200/220px with larger gaps and wrapping long labels, checkbox and text top-aligned — fields no longer crowd each other.
v0.9.20	2026-09-30	Standalone web page gains range validation: /config mirrors the Config schema min/max (recycleRetentionDays 1–3650, webPort 1–65535), returning 400 with "needs 1–3650, now 0" for out-of-range values. The most common save-failure path (cleared/out-of-range number) is now blocked with a field-level reason at all three layers: client, standalone page and host.
v0.9.19	2026-09-30	"Save failed" now says why (root cause: clearing a number box yields 0, the schema rejects it, and the host answered with an empty 500). Reproduced against the live host: out-of-range value / cleared number box / wrong type → HTTP 500 with an empty body → the client r.json() throws → only the generic "Save failed" is shown. Fixes: ① clearing a number input no longer silently becomes 0 (an empty string coerced to a number is 0, while recycleRetentionDays requires min(1)); ② pre-save per-field range validation that names the field ("recycleRetentionDays must be a number between 1 and 3650, now 0"); ③ defensive response parsing: on a non-JSON or empty body the panel shows the HTTP status and the raw snippet instead of "Save failed"; ④ the host POST is wrapped so it never returns an empty body, and returns 400 plus the field names for empty numeric values; ⑤ the standalone /config validates types and ranges so invalid pending values cannot be written; ⑥ drainPendingConfig counts failures and gives up (deleting the file) after 5 tries instead of retrying forever. npm test now runs 130 checks
v0.9.18	2026-09-30	The standalone web page can finally save config — and a read-only page no longer feels dead. Root cause (the other half of issue #12): the standalone web server has no DSH settings service, so its /config always returned writable:false; the client save() starts with if (readonly) return and the button was disabled — clicking Save did nothing at all, independent of the plugin version. Fixes: ① new shared-file protocol in lib/config-sync.js: the standalone page writes memory-eternal-config.pending.json atomically, and the DSH host applies it on activation and every 5 seconds via settings.update(patch, undefined) (skipping the optimistic-concurrency check), deleting the file on success and retrying on failure; if DSH is not running the change waits in the file for the next start. ② The standalone /config now reports writable:true, viaPending:true and the page shows a blue note explaining the sync. ③ The button is no longer disabled when read-only — clicking always explains why (🔒 Read-only + "desktop app → Settings → Memory"), so "no reaction" can no longer happen. Adds tests/config-sync.test.mjs (4 checks). npm test now runs 125 checks
v0.9.17	2026-09-30	"I updated it but the UI is still old" — fixed for good. ① The build now injects the plugin version into the client bundle and the settings page shows a "Page script vX" chip; ② when the page-script version differs from the on-disk version it says so loudly: "⚠ This page is still running an older script (page v{page} / disk v{disk}): press Ctrl+F5"; the running-vs-disk hint now also leads with "refresh first, restart the desktop app / DSH only if the host half is old". Root cause of issue #12 ("saving shows no feedback"): the page was still executing the JS loaded before the update — the host sends Cache-Control: no-store, but an already-open page never reloads its scripts, so without a refresh you keep seeing the old UI. The standalone web service on :7999 (an old process started 9/29) was restarted and now reports loaded=onDisk=latest
v0.9.16	2026-09-30	Fixed config saves silently not persisting (issue #12) + full config coverage. ① Save path: the host dsh-settings.write() requires the revision held by the panel to match describe() exactly; once stale it throws SettingsConflictError and the save fails — the plugin now re-reads the latest revision and retries once (marked ♻ in the message). ② Late volatile echo: after a successful write the patch is overlaid on the local view (reopening the panel shows the new value immediately) and the overlay is dropped once the host snapshot catches up; the host POST now re-reads after writing and returns pending with an honest note ("written; restart DSH to apply from the config file") when the echo has not landed. ③ Immediate save feedback (requested earlier): ⏳ Saving / ✅ Saved / ⚠ error now appears right next to the Save button, including whether a retry or an echo was involved, and the button label no longer wrongly says "Exporting…". ④ Full config coverage: seven fields that previously could only be edited by hand are now in the panel — the master switch enabled, single-vault vaultDir, distillation route captureProvider / captureModel, cooldown captureCooldownMs, semantic recall recallEmbedding and sessionBudgetChars; the host /config gained 6 keys and the standalone web /config gained 9. ⑤ New permanent guard tests: tests/config-coverage.test.mjs asserts every Config field is covered by the host API, the standalone web API and the settings panel (missing one turns the suite red), and tests/settings-save.test.mjs (5 checks) pins revision retry, overlay fallback and "never swallow errors". npm test now runs 121 checks
v0.9.15	2026-09-29	Server-side graph speed-ups only — rendering is identical to v0.9.13. ① On-disk graph cache keyed by fingerprint at <vault>/memory-eternal-graph-cache.json (atomic write, invalidated by fingerprint, skipped above 8MB, failures only cost a rebuild): measured cold start 422ms → 11ms at 596 cards; ② hashed bigrams + counting-sorted rare-gram index: bigramHashSet() packs each bigram into a 32-bit integer (Set<number>; capture de-duplication still uses the string version, so semantics are unchanged), and the rare-gram index is built with a counting sort (202 → 54ms); ③ MinHash + LSH candidate generation (48-dim signatures, 24 bands, unioned with the rare-gram index and then verified exactly): Jaccard calls 15,148 → 6,797 and the similarity stage 210 → 130ms. Full cold build 600 → 422ms, edges unchanged from v0.9.13 (tag 2777 / similar 4). Zero client changes — no sprites, no static layer, no skeleton degradation (the v0.9.14 client work was reverted and deprecated because it hurt the feel). Adds lib/minhash.js + tests/minhash.test.mjs (4 checks); npm test runs 108 checks
v0.9.13	2026-09-28	Seven fixes from user feedback. ① Version tracking: Plugin info now shows running (loaded by host) / on disk / npm latest side by side, with an amber warning when disk is newer ("X is installed on disk but the running process still has Y — restart to apply") plus a "Check for updates" button (server-side npm lookup, 10-minute cache, 4s timeout). ② Copy failure: the memory dialog runs inside an iframe without clipboard-write permission, so navigator.clipboard.writeText is always rejected — it now tries the Clipboard API, falls back to document.execCommand('copy'), and finally shows the text in a read-only box with a Select-all button; the iframe also gets allow="clipboard-write; clipboard-read". ③ Silent save: the result now appears right next to the Save button (⏳ Saving… / ✅ Saved / ⚠ error) and the button label changed from the wrong "Exporting…" to "Saving…". ④ Recycle bin: restore / purge / empty-all now report success and failure, and the empty state explains itself and disables the buttons. ⑤ Button feedback everywhere: the audit centre's approve / reject / delete used to swallow everything in catch {}; it now counts successes/failures and reports them (and warns when nothing is selected). ⑥ Graph zoom jank: new src/client/graph-lod.js — viewport culling (off-screen nodes/edges are skipped) + a label budget (≤120 nodes draw all; ≤300 draw 160; ≤600 draw 90; more draw 60, prioritising focus, hover, search hits and highest degree) + text-width caching + shadows off for large graphs. ⑦ Adds tests/graph-lod.test.mjs (5 checks); npm test now runs 108 checks
v0.9.12	2026-09-28	New "Report a bug" entry (🐞 in the left rail): the dialog takes your description and then either ① opens a pre-filled GitHub issue (title, body, diagnostics and the bug label already filled — one more click on Submit files it), or ② copies a ready-made AI prompt (paste it into your chat and the AI runs gh issue list to de-duplicate, files the issue and reports the URL back). Why not one-click submission? GitHub forbids anonymous issue creation, a token shipped inside the plugin would be extracted by decompiling, and OAuth/relay needs hosted infrastructure — hence "pre-fill + one click". Diagnostics come from the new /memory-eternal/api/diagnostics endpoint and are auto-redacted (home directory → ~; sk- / ghp_ / github_pat_ / api_key: → ***; truncated to 2400 chars), and the issue URL is truncated by binary search so it can never exceed the length limit (URLSearchParams inflates CJK by roughly 9x). Adds lib/feedback.js (pure functions) and tests/feedback.test.mjs (4 checks: redaction, diagnostic fields and length cap, pre-filled URL encoding/truncation, bilingual prompt); npm test now runs 103 checks
v0.9.11	2026-09-28	Fixed watchdog --reap killing itself: when you run node lib/watchdog.js --reap directly, the caller's own command line also contains watchdog.js on the same port and is not in the lock — so the reaper SIGTERMed itself, showing up as "no output at all + exit code 1" (going through dsh-memory watchdog --reap was unaffected because its command line lacks watchdog.js, which is why it only showed up on the direct path). The protected set now always includes keepPid, the current pid and the parent pid, and any other process whose command line carries --reap is skipped; output now uses fs.writeSync(2, …) so process.exit cannot truncate it. Regression test added for "self and parent must never be killed" (using a real idle child process as the target instead of a fake pid). npm test runs 98 checks
v0.9.10	2026-09-28	Multi-vault + watchdog wrap-up, plus graph frame-time work. ① Multi-vault (#10): vaultProfiles / activeVault finally have a settings-page form (add/remove vaults, set directories, pick the active one, and see the effective directory / source / workspace); vault resolution moved into lib/vault-resolve.js and is shared by the host and CLI / MCP / hooks / standalone web / sweep, with new vaultProfiles[].match.workspace auto-selection by project directory (longest prefix wins, opt-in, still single-vault by default); memory_recall gained a scope argument (vault name / path prefix / all for cross-vault aggregation) and cross-vault results are labelled with their source vault. ② Watchdog (#11): the lock is now per-port (several ports/vaults no longer overwrite each other) and dsh-memory watchdog --reap collects orphans left by earlier versions (only same-port processes that are not in the lock; the live registered instance and user-started watchdogs are left alone). ③ Graph frame time: edges are bucketed by colour + alpha + width and stroked with a single Path2D per bucket (2.7k edges: 2.7k draw calls → a handful), and node radial gradients are cached per colour|radius and reused via translate (593/frame → essentially none after the first frame). ④ New tests/vault-resolve.test.mjs (7 checks), watchdog tests grown to 7; npm test now runs 98 checks
v0.9.9	2026-09-28	Fixed the sidebar "Memory" entry being pushed out of view (issue #2): the slot host comes in two shapes — the shell's own horizontal row (.footerActions) and a div[data-slot="sidebar.footer.action"] that other plugins (dsh-diff-approval, dsh-footer-order, …) turn into flex-direction: column. The old CSS applied flex-wrap: wrap plus .me-footer { flex: 1 1 100% } to both: inside a column container wrap means "new column" and basis: 100% means 100% height, so the button was pushed into an overflow column and clipped to a sliver. flex-wrap is now scoped to the horizontal .footerActions only and .me-footer uses the direction-agnostic flex: 0 0 auto, guarded by a new SSR assertion. Also: bilingual READMEs gained a "Community & Support" section (group QR + WeChat pay/reward codes), plus SPONSOR.md and .github/FUNDING.yml
v0.9.8	2026-09-28	Backlog burn-down release. ① Capture reliability (issues #3 / #4): distillation no longer hardcodes providers[0] — new captureProvider / captureModel config plus ordered fallback across providers; adapter errors carried by the stream's terminal finish block (missing credential, rate limit, …) are now visible failures (code/message in the capture log, red health state, raw-card fallback annotated with the cause) instead of a vague "distillation produced nothing"; the stall detector moved from cumulative-counter deltas to a sliding window (alerts only when turns keep starting with zero finishes inside 20 min), fixing both false alarms and misses. ② Independent-process vault consistency (dsh-memory-eternal#5): CLI / MCP / hooks / standalone web / sweep now share the host's resolution order (MEMORY_VAULT_DIR → activeVault → vaultDir → default), so switching vaults no longer splits your memory in two. ③ Watchdog singleton lock (dsh-memory-eternal#6): a new process exits when a live watchdog already owns the port, and stale pid records are cleaned, ending the orphan pile-up on every restart. ④ Graph performance: tag edges became "Top-K by shared tags" (measured 24,193 → 2,763 edges at 593 cards), similarity switched to a rare-bigram inverted index (cold 948 → ~520ms, warm 2-4ms), payload 3,531KB → 579KB, and the client dim pass now uses an adjacency set (52ms → <1ms per frame). ⑤ Five new test files (routing/finish, stall window, vault resolution, watchdog lock, graph perf baseline); npm test now covers 85 checks
v0.9.7	2026-09-28	Cards open fully expanded (and can be collapsed) + new body typography: the reader no longer folds to 300px by default — it opens showing the whole card, with a one-click collapse for long ones (and an "Expand ⌄" hint when folded). Frontmatter is lifted out of the body into colored chips (kind / audit status / tags / dates / source; kind reuses the graph palette) instead of leaking raw YAML. Body rendering is rewritten (src/client/markdown.js, shared by the host and the standalone web page): headings get an emoji + color group by meaning (conclusion ✅green / risk ⚠️amber / root cause 🔍blue / plan 🛠️purple / example 💡pink), consecutive - lines become a real <ul>, 1. becomes an <ol> with numbered circles, - [x] becomes a checkbox, > becomes an emoji callout, --- becomes a divider, short key: value lines get an aligned style, and **bold** gets a highlighter underline; fenced code is protected so # / - inside it are no longer re-typeset. Adds 7 typography unit tests (XSS escaping, link-protocol allowlist, prose colons must not be misread as key/value). Also carries the PR #8 port: Electron-host subprocess node resolution (lib/node-bin.js, fixing "standalone web server never starts → sidebar popup white screen") and API routes registered after webServer is ready (fixing "settings page stuck loading")
v0.9.6	2026-09-28	Fixed the full-screen overlay covering the desktop shell's window controls (clicking Close quit the whole shell): the desktop shell draws minimize / maximize / close at the top of the page and, per the DSH convention, marks <html> with data-windows-titlebar and publishes --dsh-windows-titlebar-height (both confirmed present inside the official desktop build's app.asar); the memory overlay used inset:0 plus an inline 100vh, covering that whole strip, and our own top-right "×" landed exactly on the shell's window Close — one click quit the entire desktop shell. The full-screen state now uses the .me-modal-full class (no inline 100vh, which CSS cannot override), and under [data-windows-titlebar] the .me-overlay-top / .me-overlay / .me-modal-full surfaces are pushed down by --dsh-windows-titlebar-height with height calc(100vh - that). Browsers without the attribute get a 0px offset, so behaviour is unchanged there. The export-graph full-screen preview also moved from 100vh to 100% so it cannot overflow the shortened overlay. Adds 1 render regression asserting all three avoidance rules ship in the bundle
v0.9.5	2026-09-28	Fixed the plugin failing to mount under the official desktop build (schemastery 3.18.4): the desktop profile resolves schemastery 3.18.4, which turns every volatile Config field into a cosmokit live reference ({ get(): snapshot } branded with Symbol.for('cosmokit.volatile.write')), while 0.9.4 treated those fields as plain values — cfg.vaultDir.trim() threw TypeError: cfg.vaultDir.trim is not a function and the plugin died inside apply (the exact desktop error). The plugin now deep-unwraps references per the cosmokit protocol (snapshot each reference, recurse through arrays/objects), so settings.get() only ever exposes plain values; each read unwraps afresh, so in-place volatile hot updates stay immediately visible, and JSON.stringify no longer writes {} into the shared config file. Adds 2 regressions (the real 3.18.4 reference shape + nested arrays/objects) and verifies apply under the desktop build's real module-resolution layout (schemastery 3.18.4 + real cosmokit)
v0.9.4	2026-09-28	DSH v0.1.7-rc.2 support (fixes "the plugin goes dark after the upgrade"): (1) The settings service was replaced — in 0.1.7 ctx.settings is a forms-only API (describe/update/replace/mutate/configure) with no register(); the old ctx.settings.register(...) call throws a TypeError, so the plugin never mounts (memory page gone, auto-capture dead). A cross-version shim keeps register for older hosts while new hosts read the live config reference the loader passes in, subscribe to loader/volatile-update for hot updates, and write back through settings.update(entry id, patch, revision) (revision conflicts map to HTTP 409); it also registers configure({ auto:false }) so the host does not generate a duplicate form page. (2) Volatile field marking moved after construction: no more chained .volatile() — that method only exists in schemastery ≥3.18.4, while the plugin actually resolves 3.18.1 at the profile level, so the chain threw volatile is not a function at import time and killed the plugin. (3) Dead client references removed: inject no longer lists settingsScope, which disappeared together with @deepseek-ai/dsh-client-runtime (0.1.7 has no such service, so the client plugin waited forever and never activated — taking the memory page and the sidebar button with it); dsh.client.inject dropped its non-existent package names. (4) New tests/settings-compat.test.mjs (8 cases: 0.1.7-shaped mount / live-reference reads / volatile hot update / write-back and revision conflict / legacy register path / field volatile markers) wired into npm test. Verified by booting a real DSH v0.1.7-rc.2: the host plugin mounts (config written to memory-eternal-config.json) and the client bundle composes into the boot graph (plugins/??memory-eternal/client.js)
v0.9.3	2026-09-12	Fixed the blank memory page ("no UI"): v0.9.1 placed the infinite-scroll useEffect before the loadCards definition — React evaluates the dependency array during render, hitting a TDZ error (Cannot access 'loadCards' before initialization) that crashed the whole memory page on first render. It now comes after the definition. Added a render smoke test (tests/ssr.test.mjs: renders the built bundle through all four entry points — cards / settings / usage / graph — wired into npm test) so "green build, blank page" bugs cannot ship again
v0.9.2	2026-09-12	Log visibility + false-alarm fix: the capture log is now persisted to disk (~/.dsh/memory-eternal-capture.jsonl, last 500 entries) — the standalone web page no longer says "log only available inside the DSH host" but shows the records, and history survives a host restart (read back on boot); fixed stall-detector false alarms flooding the panel (mismatched turn counters are normal: sub-agent turns and cancelled turns emit no turn-stopping; it now requires "turns running and no capture activity for 15 minutes" to alert once, and the alarm clears as soon as activity resumes — no more one fail line every 10 minutes permanently turning health red); when distillation returns nothing / unparseable JSON the content is saved as a raw card instead of being dropped
v0.9.1	2026-09-12	Performance & UX batch: memory page loads lazily per view (opening Settings no longer drags in 250KB of cards + 890KB of graph; first paint needs ~0.5KB); graph generation 3232ms → 43ms (bigram sets computed once + lossless mathematical-bound pruning + tag inverted index) plus a server-side cache with fingerprint invalidation (write/audit/delete rebuilds immediately, no stale reads); site-wide brotli q5 (graph 877KB → 54KB, 34% smaller than gzip; cards 246KB → 83KB → 20KB with br) and large responses compressed with both br and gzip keeping the smaller, so no response is ever worse than gzip; /cards gained real pagination (offset/limit/total + server-side title sort, globally ordered across pages) loading 100 cards first and fetching more on scroll; the "Agent cards" panel in Usage/Today filters by agent server-side (30 rows); audit red line: the plugin injects "never call approve/reject on the user's behalf" into the agent context; the watchdog now reaps redundant instances on fallback ports (a duplicate web process holding 76MB was reaped in testing)
v0.9.0	2026-09-12	Fixed silent auto-capture death (6 days without a single card): after a DSH upgrade the session event API moved from events to ownEvents()/snapshotEvents(); the listener read nothing, returned silently and left no trace — now it tries ownEvents() → snapshotEvents() → events and logs a per-session attach entry. Daily quota now counts real writes only (previously every "not worth saving" verdict burned quota, exhausting the 24h budget after 40 attempts); captureCooldownMs actually applies; a single capture input is tail-capped at 20k chars. New capture log panel and active failure alerts (red strip + agent-context injection + turn-stopping stall sweep). Removed the automatic "Daily review" digest (timer / /todayBrief route / UI button).
v0.8.0	2026-09-06	Full UI i18n: follows the DSH system language live (zh/en) across cards, graph, audit center, config panel and new-card templates; SQLite EBUSY fix for the Windows test suite
v0.3.1	2026-09-03	Knowledge graph loads on demand (fetch on open, unregister on close)
v0.3.0	2026-09-03	SQLite storage layer (node:sqlite, zero deps) + DB-level audit guard enforceAudit() + audit_log table + automatic .md migration
v0.2.0	2026-09-03	Recall with body text (search() returns the excerpt field)
v0.1.2	2026-09-03	Audit guard hardening (writeCard defaults to pending, mergeCards goes through audit)
v0.1.1	2026-09-03	Recall body fix (excerpt field + body logic)
v0.1.0	2026-08-31	First release: auto capture + auto recall + graphical knowledge base + knowledge graph + audit center + recycle bin

---

💬 Community & Support

Questions, feature requests, or just want to talk about DSH plugin development — join the group chat:

If this plugin saved you some time, you can buy the author a coffee (either code works):

WeChat Pay	WeChat Reward

Donations are entirely optional and change nothing about the plugin; it stays open source and free (MIT). Picking the first custom entry in the repo's Sponsor dropdown lands on SPONSOR.md, which shows the same codes.

---

📄 License

MIT

---

Make your AI truly remember: dialogue auto-captured, knowledge at your fingertips. ⭐ Star it if you like, Let's make AI not forget.

中文 README: README.md