dsh-memories
DeepSeek Harness 双账本跨会话记忆：自动提取的长期事实（MEMORY.md）以及持续更新的项目进度账本（PROGRESS.md），并在每个新会话中调取。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add dsh-memories@0.2.1Memories 以 dsh-memories 发布，当前版本为 0.2.1。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
Dual-ledger cross-session memory for DeepSeek Harness.
Inspired by the memory pipeline of OpenAI's open-source Codex agent, dsh-memories gives every new session in a project two living ledgers:
| Ledger | File | Answers |
|---|---|---|
| Long-term facts | .dsh/memories/MEMORY.md | "What are this project's conventions, the user's preferences, past pitfalls?" |
| Project progress | .dsh/memories/PROGRESS.md | "Where are we? What's done? What's in progress? What's next?" |
Both are maintained automatically and recalled into the system context of every new conversation — no re-briefing required.
.bak backups).pinned.remember / update_progress tools let the agent write proactively; /remember, /progress, /memories give humans direct control.A DeepSeek Harness deployment that provides the standard host-plane services:
fs · llm · sessionQuery · systemPrompt · tools · commands · sandboxPolicy
(all ship with the default dsh-base bundle; tested on dsh 0.1.0-rc.9)
dsh plugin add (recommended)dsh plugin --profile web add dsh-memories
The plugin manager reads this repo's bundled `cordis.patch.yml) and wires everything for you. Needs a release that ships the bundle manifest -- the current npm 0.1.2 predates it, so until the next publish, use Option A.
cd ~/.dsh/profiles/web # or whichever profile you run
npm install <git-url-or-tarball> # places dsh-memories into ./node_modules
Then append a row to ~/.dsh/profiles/web/cordis.patch.yml. Reference the entry file by relative path — the proven pattern in pnpm-managed profiles (bare package names in patch rows are not reliably resolved):
- insert:
- id: dsh-memories
name: './node_modules/dsh-memories/lib/index.js'
Alternatively, register it npm-natively: add "dsh-memories": "*" to the profile package.json dependencies, add "dsh-memories" to dsh.profile.bundles, run pnpm install, and skip the patch row entirely.
Copy this repository folder into ~/.dsh/profiles/web/node_modules/dsh-memories, then add the same patch row as above.
Restart DSH once. Done — the plugin loads at boot and works silently afterwards.
If you previously ran this capability as a dynamic plugin (
cordis_define/mem-*), remove it before restart to avoid duplicate tool registrations.
You normally do nothing. For direct control:
| Command / Tool | Effect |
|---|---|
/remember <fact> | Append a fact to the current project's ledger and trigger consolidation |
/memories | Status: processed count, draft files, ledger previews, last error |
/memories rescan | Force a fresh extraction pass now |
/memories reset | Clear the "processed" registry so recent sessions get re-read |
/progress | Show PROGRESS.md |
/progress <note> | Append a progress note manually |
model tool remember(fact, category?, pinned?) | The agent saves a durable fact mid-conversation |
model tool update_progress(completed?, doing?, next?) | The agent updates the progress ledger at milestones |
<your-project>/.dsh/memories/
├── MEMORY.md # consolidated long-term facts (4 sections)
├── MEMORY.md.bak # previous version, kept on every rewrite
├── PROGRESS.md # project progress (# 已完成 / ## 进行中 / ## 下一步)
├── PROGRESS.md.bak
└── raw/ # unconsolidated drafts (_manual.md, _progress.md, per-session notes)
Everything is plain Markdown — open it, edit it, diff it.
new session ──▶ scan (same-workspace sessions, ≤14 days, ≤2/run)
│
▼
extract (one LLM call per session)
├─ facts[] → raw/<session>.md
└─ progress{} → raw/_progress.md
│
▼
consolidate (one LLM call per ledger)
├─ MEMORY.md ← merge + dedupe + decay(30d) + sections
└─ PROGRESS.md ← 已完成 / 进行中 / 下一步 + dates
│
▼
next session ◀── recall section injected into system prompt
The extraction prompt enforces a strict schema (JSON array / NONE), caps entry length, forbids secrets, and skips trivial transcripts (<400 chars). Consolidation prompts enforce section formats and size budgets (≤100 lines for facts, ≤60 for progress).
~/.codex/memories) is planned via a globalDir setting.MIT