dsh-remember

中文

Cross-session long-term memory for the DeepSeek Harness(dsh).

While you're away, it looks back at your recent sessions and distills the useful experience into a set of plain Markdown files. Every conversation after that starts with a summary of those memories, and the model greps the originals when it needs details.

No vector database, no external service — memories are just files under ~/.dsh-memory/ that you can open, search, and manage with git.

Features

• Automatic extraction — idle sessions are scanned, redacted, and distilled in the background
• Automatic consolidation — new experience is merged incrementally; stale, never-cited entries are pruned
• Automatic recall — every prompt carries a memory summary (configurable token cap); the model searches on demand
• Readable and controllable — a sidebar status panel, a memory viewer, and a settings page; pause anytime
• Translatable — one-click translation of the memory files with incremental caching
• Portable — gzip bundle export/import with content-hash dedup

Screenshots

The memory entry at the bottom of the sidebar opens the status panel.

The memory viewer: overview, summary, index, and records.

Install

From npm:

dsh plugin --profile <name> add dsh-remember

From source:

git clone https://github.com/Arborsm/dsh-remember.git
cd dsh-remember
pnpm install && pnpm build
dsh plugin --profile <name> add .

Restart dsh web after installing. During development you can skip installation and load straight from the workspace (run from the dsh checkout root):

pnpm dsh web --patch <path/to>/dsh-remember/cordis.dev.yml

Note: --patch must come before --port, otherwise commander fails with unknown option '--patch'.

Configuration

Every field has a default; the plugin works out of the box. Override via cordis.yml or the config: block of a dev overlay. Field definitions live in src/config.ts.

Commonly tuned:

Field	Default	Description
enabled	true	Master switch
generateMemories	true	Run extraction and consolidation
useMemories	true	Inject the memory summary into prompts
extractModel / consolidationModel	''	provider/model; empty = deployment default
workspaceDir / dbPath	''	Empty = under ~/.dsh-memory/
summaryTokenLimit	2500	Token cap for the injected summary
scanIntervalMinutes	15	Background scan interval
minRolloutIdleHours	1	Idle time before a session becomes extractable
maxUnusedDays	30	Retention for never-cited memories

There are 23 fields in total (scheduling, retries, concurrency, consolidator tool deny-list, …) — see src/config.ts. All of them are editable on the settings page.

Memory files

~/.dsh-memory/
├── memories/                  # git workspace
│   ├── MEMORY.md              # index of all memories
│   ├── memory_summary.md      # the summary injected into prompts
│   ├── rollout_summaries/     # source evidence for each memory
│   ├── skills/                # distilled skills
│   └── …                      # translations, ad-hoc notes, …
└── memories.sqlite            # extraction queue and job state

Development

pnpm build        # build (tsdown, host + browser bundles)
pnpm typecheck    # tsc --noEmit
pnpm test         # node --test
pnpm shots        # regenerate the screenshots above (fixtures + headless Chrome)

Repository layout, conventions, and debugging notes: AGENT.md.