MIT, zero-runtime-dependency local-first memory system (Node ≥ 22, built-in `node:sqlite` FTS5) that DeepSeek Harness mounts as a stdio MCP server, so decisions and gotchas survive as an immutable append-only log you can query `as-of` any point in time.
DSH integration
Native runtime
Author-claimed
Safety audit
Unaudited
Last verified
2026-11-15
License
MIT
01What can it help you accomplish?
Keep one memory that DeepSeek Harness, Codex CLI and Claude Code all read and write, so context survives across sessions
An append-only truth layer at `~/.memo/items/<scope>.jsonl`, queried through `memory_search` with BM25 plus an optional semantic lane fused by RRF, and inspectable `as-of` any past point in time
DSH users who keep re-explaining the same project facts, decisions and gotchas to every fresh session
Govern an accumulated memory instead of letting duplicates, conflicts and stale facts rot
Dedupe scans, dedupe-cluster arbitration, supersede of old values and tombstone deletes, plus a hygiene page showing pending clusters and a 30-day trend
Long-lived local setups where near-duplicate or contradictory notes pile up and nobody trusts the memory anymore
Look at what the agent actually remembered today without touching the data
The nine-page memory center on http://127.0.0.1:7788 — today, graph, search, hygiene and more — read-only unless you start it with `--write`
Anyone who wants to verify the harness memory before relying on it in a real project
02How to install into DeepSeek Harness
Prerequisites
- Node ≥ 22 — sedimem has zero runtime dependencies and uses Node's built-in `node:sqlite` FTS5.
- A local checkout of the repo (`git clone https://github.com/Old-Friends-Gathering/sedimem`).
- A DSH profile whose `~/.dsh/profiles/<profile>/cordis.patch.yml` you can edit.
Installation steps
- 01
Clone and walk through the README 30-second quickstart: `git clone https://github.com/Old-Friends-Gathering/sedimem && cd sedimem`, then run the quickstart lines `memo.js write --kind fact`, `memo.js search` and `memo.js serve --port 7788` with the Node binary from `command -v node` and `V=versions/$(cat versions/CURRENT)`.
$ git clone https://github.com/Old-Friends-Gathering/sedimem && cd sedimem
- 02
Insert the DSH block into `~/.dsh/profiles/<profile>/cordis.patch.yml`: `id: memory-sedimem`, `name: '@deepseek-ai/dsh-mcp-client'`, `config.serverName: memo`, `transport: stdio`, `command: node`, `args: ['/ABS/PATH/versions/v1.1/src/mcp/server.js']`, `env: { MEMO_HOME: '$HOME/.memo' }`.
- 03
Before the first read or write, call `memory_scope` in DSH to align the partition — DSH is a long-running process.
- 04
Optional: point Codex CLI (`~/.codex/config.toml`, `[mcp_servers.memo]`) and Claude Code (`claude mcp add --scope user memo ...`) at the same `server.js` and `MEMO_HOME`.
Verify the integration
- Run `bash versions/v1.1/configs/verify-three-clients.sh` — the README says to expect `PASS`.
- Confirm the tool surface: 10 `memory_*` tools plus 5 `memory_session_*` tools, identical across the three clients.
- Two read-only commands need no setup: `memo doctor` and `memo stats`.
03DSH integration and capability boundaries
Native DSH MCP mount: one `@deepseek-ai/dsh-mcp-client` block inserted into `~/.dsh/profiles/<profile>/cordis.patch.yml` (serverName memo, stdio, `node .../versions/v1.1/src/mcp/server.js`, `MEMO_HOME=$HOME/.memo`)
Append-only truth layer with four write gates
A fact / decision / gotcha written via `memory_write` or `memo write`→A new version appended to `~/.memo/items/<scope>.jsonl`, with the old version still readable `as-of`
keys, internal domains and private paths are redacted to `[REDACTED:x]` before anything reaches diska duplicate or conflicting slot is marked superseded / disputed instead of overwriting the old factHybrid retrieval (BM25 + optional semantic lane)
A query string via `memory_search`→Results fused with RRF (k=60) that carry their basis, explicitly `degraded` when the semantic lane is unavailable
Read-only memory center (nine pages)
`memo serve --port 7788`→Today / graph / search / hygiene / edit / import / session / handoff / settings pages at http://127.0.0.1:7788
read-only by default — editing needs `--write`, a token and If-MatchTask library for exploratory work (`session:` scope)
`memo session start / checkpoint / status / close / sweep`→Only decision / gotcha entries promoted to the main store with a `from-session:<id>` tag, plus a six-section handoff document at `docs/handoffs/<id>.md`
`session:*` scopes reject `fact` writeseverything not promoted is discarded together with the temporary libraryOptional LLM derivation layer
`memo derive` / `memo llm` pointed at an OpenAI-compatible endpoint→Titles, summaries and tags in the derived sidecar `derived/<scopeKey>.jsonl`
when enabled, title / summary / tag generation requests are sent to the provider you configured; the raw memory text stays local
04Who is it for? When not to use it?
Good for
- DSH users who keep re-explaining the same project facts, decisions and gotchas to every fresh session
- Long-lived local setups where near-duplicate or contradictory notes pile up and nobody trusts the memory anymore
- Anyone who wants to verify the harness memory before relying on it in a real project
Not for
- Node ≥ 22 is required: sedimem has zero runtime dependencies and relies on Node's built-in `node:sqlite` FTS5, so an older Node cannot run it.
- A `scope` is a classification, not a permission boundary — sharing a memory between people would need a separate deployment with authentication, which the project deliberately does not ship.
05Compatibility, maintenance and safety notes
- Node ≥ 22 is required: sedimem has zero runtime dependencies and relies on Node's built-in `node:sqlite` FTS5, so an older Node cannot run it.
- Local-first by default, but enabling the optional LLM layer sends title / summary / tag generation requests to the provider you configure; the truth layer itself never takes part in those requests.
- A `scope` is a classification, not a permission boundary — sharing a memory between people would need a separate deployment with authentication, which the project deliberately does not ship.
- No cloud sync and no multi-user permissions by design; all data lives in `~/.memo/`, so deleting the directory is the only erase path and nothing is backed up.
MIT · no published release · last push 2026-11-13
06Frequently asked questions
How do I mount sedimem in DeepSeek Harness?
Insert one block into `~/.dsh/profiles/<profile>/cordis.patch.yml`: `id: memory-sedimem`, `name: '@deepseek-ai/dsh-mcp-client'`, `serverName: memo`, `transport: stdio`, `command: node`, `args: ['/ABS/PATH/versions/v1.1/src/mcp/server.js']`, `env: { MEMO_HOME: '$HOME/.memo' }`. Because DSH is a long-running process, call `memory_scope` first to align the partition.
Can DSH, Codex CLI and Claude Code share a single memory?
Yes — configure one stdio MCP entry per client, all pointing at the same `server.js` and the same `MEMO_HOME`. The tool surface is identical everywhere: 10 `memory_*` tools plus 5 `memory_session_*` tools, and `bash versions/v1.1/configs/verify-three-clients.sh` should print `PASS`.
Does DSH need an LLM or an API key for sedimem to work?
No. `provider=none` is the default and the whole LLM layer simply does not exist — search, brief, hygiene and the graph run at zero LLM cost. If you enable it, titles, summaries and tags are written only to the derived sidecar, and the truth layer bytes never change.
Where does the memory live, and how do I erase it?
Everything sits in `~/.memo/`, with the truth layer at `~/.memo/items/<scope>.jsonl` — append-only, never rewritten in place, so past states stay queryable `as-of`. Deleting the directory clears it completely; there is no cloud component and no telemetry.
What happens when two memories contradict each other?
Writes run the four gates in order, and a conflicting value for the same slot becomes `superseded` or `disputed` and can be handed to the agent for arbitration through `memory_curate`; `memory_forget` deletes by writing a tombstone by default.
07Related DSH workflows
dsh-univer-office
by dream-num
Give DeepSeek Harness a real office environment. Univer Office Plugin brings spreadsheets, docs, slides, canvases, relational tables, and more into one runtime — with connected data, validation, versioned changes, and isolated worktrees for multi-agent collaboration.
clearai-dsh
by clearailhc
Brings an evidence-checked epistemic loop to DSH: research sessions settle into a browsable domain ontology with evidence chains and support levels; contradictory readings are surfaced for human adjudication.
dsh-fish-memory
by yuuyuko-uu
Stores time-stamped, decaying per-session memories, recalls relevant items, and optionally distills imported local archives through the configured DSH model.
dsh-trading
by zhu1090093659
Agent-native trading terminal built on DeepSeek Harness. Crypto, US, CN and HK in one three-column GUI, 19+ hot-swappable connectors, dry-run by default with human approval on every live order. BYOK, no data redistribution.
08Data and sources
- 🔌 **三端共用一个记忆**:DSH / Codex CLI / Claude Code 连同一个 stdio MCP server、同一个分库
# DSH:~/.dsh/profiles/<profile>/cordis.patch.yml - insert: - id: memory-sedimem name: '@deepseek-ai/dsh-mcp-cl…
serverName: memo transport: stdio command: node args: ['/ABS/PATH/versions/v1.1/src/mcp/server.j…
三端各配一条 stdio MCP,指向同一个 `server.js` 和同一个 `MEMO_HOME`。常驻进程(DSH)记得先 `memory_scope` 对齐分库。
工具面:10 个 `memory_*` + 5 个 `memory_session_*`,三端完全一致。
This page is generated from the project’s public documentation, repository metadata and a structured parse of DSH Plugins; last verified on 2026-11-15. Found an error? Submit a correction.
Best DeepSeek Harness Plugins
Twelve plugins worth installing first — picked from the whole catalog, across every category.
