dsh-unified-memory
以 Claude Code 的格式为 DeepSeek Harness 提供长期记忆：可使用私有存储、同时读取 Claude Code 记忆的叠加层，或与 Claude Code 共享的单一存储，并为每个文件记录来源信息。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:mattcarvercom/dsh-unified-memory#164ef831d94b56c937d8b95a1e25279794a11fc0Unified Memory 以 dsh-unified-memory 发布，当前版本为 1.0.0。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。
Long-term memory for DeepSeek Harness (dsh), stored as plain markdown files in Claude Code's memory format. Run it as dsh's own private store, as a private store that also reads Claude Code's memories, or as one store shared with Claude Code so both agents read and write the same memories.
Every file dsh creates says so in its frontmatter, which makes switching modes a mechanical migration and lets you remove everything dsh ever wrote with one command.
| Mode | dsh writes to | dsh reads and searches | Claude Code sees dsh's memories |
|---|---|---|---|
dsh | the dsh store (~/.dsh/memory) | the dsh store | no |
overlay | the dsh store | the dsh store, plus Claude Code's memories (read-only) | no |
shared | Claude Code's memory folder | Claude Code's memory folder | yes |
dsh plugin --profile web add github:mattcarvercom/dsh-unified-memory#v1.0.0
Restart dsh web afterwards. The plugin starts in dsh mode (its own store in ~/.dsh/memory); see Modes to share memories with Claude Code.
To work on the plugin itself, install from a checkout instead. dsh then loads it from outside the profile, so link the checkout's node_modules to the runtime's for its @deepseek-ai/* imports to resolve:
git clone https://github.com/mattcarvercom/dsh-unified-memory
ln -s ~/.dsh/profiles/node_modules dsh-unified-memory/node_modules
dsh plugin --profile web add ./dsh-unified-memory
| dsh | Tested with DeepSeek Harness 0.1.7-rc.1, web profile |
| Node | Whatever dsh requires (22.19+ or 24+), which includes the built-in node:sqlite used for search; tested on Node 24 |
| Claude Code | Key and folder rules verified against Claude Code 2.1.282; only needed for overlay and shared modes |
| Platforms | Linux and macOS tested paths; Windows paths follow the same key rule but are untested |
| Dependencies | None beyond the dsh runtime (@deepseek-ai/dsh-tools, @deepseek-ai/schemastery) |
| Network | None. Nothing leaves your machine |
| Files read | The dsh store (~/.dsh/memory) and, in overlay and shared modes, ~/.claude/projects/*/memory |
| Files written | The dsh store, plus Claude Code's memory folders in shared mode only; search indexes in ~/.cache/dsh-unified-memory |
| Processes | Runs git rev-parse to find a session's repository root |
Memories are plain text on disk. Do not store secrets in them; the injected prompt tells the agent the same.
Override the memory row in your profile's cordis.patch.yml. The override replaces the row's config wholesale, so restate every field you set.
- id: memory
config:
mode: shared
promptMode: feedback
| Key | Default | Meaning |
|---|---|---|
mode | unset | dsh, overlay or shared. Unset follows whatever mem mode last recorded |
dshRoot | $DSH_HOME/memory, else ~/.dsh/memory | The dsh store |
claudeRoot | $CLAUDE_CONFIG_DIR or ~/.claude | Claude Code's home |
projectAliases | {} | Project key to canonical key, so several checkouts share one folder |
promptMode | full | full injects each folder's MEMORY.md and a file list; feedback injects only the standing feedback rules and relies on search |
includeInPrompt | true | Inject the memory section into the system prompt |
maxIndexLines / maxIndexBytes | 200 / 25000 | How much of MEMORY.md is injected |
maxList | 200 | Cap on files listed in full mode |
searchLimit | 5 | Default hits returned by search |
indexDir | $XDG_CACHE_HOME/dsh-unified-memory | Search index cache |
minResyncIntervalMs | 60000 | Minimum time between automatic rescans of a folder for search; dsh's own writes reindex immediately |
Setting mode never moves files. The files on disk are in the mode recorded in <dshRoot>/state.json. When the configured mode differs, the plugin stays readable but refuses writes and tells the agent to have you run mem mode.
A fresh install adopts the configured mode (or dsh) on first use, since there is nothing to migrate.
mem modemem mode # print the current mode
mem mode shared --dry-run # show exactly what would move
mem mode shared # back up, migrate, rebuild search indexes
| From | To | What moves |
|---|---|---|
dsh | overlay | nothing |
overlay | dsh | nothing |
dsh or overlay | shared | dsh's memories move into Claude Code's folders, with their MEMORY.md lines |
shared | overlay | memories dsh created move out to the dsh store |
shared | dsh | as above, plus copies of Claude Code's memories, so dsh keeps what it knew; --clean skips the copies |
Details that keep this safe:
<dshRoot>/backups/<timestamp>/. Only the memory folders are copied, never Claude Code's session transcripts.-dsh (or -claude) suffix, and the dry run shows it.--prefer dsh (dsh's version replaces the original's body and description) or --prefer claude. A copy whose original you deleted is dropped unless you pass --restore-deleted.Files dsh creates carry provenance under metadata::
---
name: deploy-notes
description: How myapp deploys
metadata:
type: project
origin: dsh
originSessionId: session-7c1e4b2a-9d3f-4a6b-8e5c-2f1a0b9c8d7e
modified: 2026-09-24T18:02:11.000Z
---
| Field | Meaning |
|---|---|
origin | Who created the file. Absent means Claude Code in Claude Code's folders, and dsh in the dsh store |
originSessionId | The dsh session that created it (Claude Code writes the same field for its own files) |
updatedBy | Added when dsh edits a file it did not create; origin never changes |
copiedFrom, copiedAt | On copies taken when leaving shared mode |
mem list --origin dsh # everything dsh created
mem list --updated-by dsh # Claude Code memories dsh edited
mem purge --origin dsh --dry-run
mem purge --origin dsh # back up, then delete dsh's files and their index lines
Purge never removes files dsh only edited; it lists them. When dsh deletes a memory it did not create, a copy goes to <dshRoot>/trash/.
MEMORY.md index with one - [Title](file.md) — hook line per memory.type: at the top level keeps it there.MEMORY.md first. Claude Code takes no lock, so the plugin reads the index immediately before each change to keep a concurrent edit from being lost.user, feedback, project, reference.Claude Code keys memory by the git repository root of the directory it was launched from (a worktree maps to its main checkout), or the directory itself outside a repository. The key replaces every non-alphanumeric character with -, so /home/me/dev becomes -home-me-dev; keys over 200 characters are cut and given a hash suffix.
dsh sessions often start deeper than where Claude Code was launched, so the plugin walks from the session's root up to the nearest ancestor that already has a memory folder on either side, stopping at your home directory. Without a match, the root's own key is used. projectAliases apply before the walk.
memory action=search and mem search use a SQLite FTS5 index (built into Node's node:sqlite, no native dependency) with Porter stemming and bm25 ranking, name and description weighted above the body. The markdown files are the only source of truth: the index lives in indexDir, is updated incrementally, and can be deleted at any time.
With promptMode: feedback, only the standing feedback rules are preloaded (about 1k tokens for a hundred memories) and the agent searches for everything else per task, which suits models with small context windows.
The memory tool offers add, update, delete, search, list and read. add refuses a memory whose name or description closely matches an existing one and lists the matches, so the agent updates one of them instead; force: true saves anyway. The model decides when to save; the injected prompt tells it what is worth keeping.
mem referencemem status [--cwd DIR]
mem search <keywords...> [-n N] [-o OFFSET] [-t TYPE] [--cwd DIR] [--json]
mem list [--origin dsh|claude] [--updated-by dsh] [--json]
mem mode [dsh|overlay|shared] [--dry-run] [--clean] [--prefer dsh|claude] [--restore-deleted]
mem purge --origin dsh [--dry-run]
mem move-legacy [--cwd DIR] [--dry-run]
mem reindex [--rebuild]
Common options: --dsh-root, --claude-root, and --alias RAW=KEY (repeatable). $DSH_MEMORY_ROOT (else $DSH_HOME/memory) and $CLAUDE_CONFIG_DIR set the defaults.
memoryRoot, scope and importClaudeMemory settings are gone (still accepted, and ignored with a warning). Use mode, dshRoot and claudeRoot.dsh mode. Run mem mode <mode> to move to another mode.~/.dsh/memory/*.md) stay readable but read-only; mem move-legacy --cwd <project> folds them into a project.import-claude tool action is gone: overlay mode reads Claude Code's memories in place, and mem mode dsh from shared takes copies.~/.cache/deepseek-harness-memory to ~/.cache/dsh-unified-memory; the old folder can be deleted.Anything else that reads or writes a folder shared with dsh-unified-memory and Claude Code should:
.md file per memory, with YAML frontmatter holding at least name and description, and a type either as top-level type: or as metadata.type.- [Title](file.md) — hook line per memory in MEMORY.md, change only its own lines, and re-read the file immediately before writing it.metadata.origin: <tool> and files it edits with metadata.updatedBy: <tool>, never changing another tool's origin.ln -s ~/.dsh/profiles/node_modules node_modules # runtime imports, for the plugin tests
npm test
dsh-unified-memory began as a fork of deepseek-harness-memory by jzc (MIT). Its frontmatter parsing and a few storage helpers still derive from that work; the storage modes, provenance tracking, migrations, full-text search, Claude Code key resolution, and the mem CLI are new.