MIT-licensed, TypeScript DSH host plugin that keeps long-term memory in local JSONL with dual-channel (lexical + local llama-server) retrieval, per-turn recall, hard rules, rule hook scripts and a five-tab panel — no vector service required, no network requests unless you configure one.
DSH integration
Native runtime
Author-claimed
Safety audit
Unaudited
Last verified
2026-11-18
License
MIT
01What can it help you accomplish?
Keep a DSH session connected to the conclusions of earlier sessions
Memories stored as local JSONL, a project / global digest injected at the first step of a new session, and per-turn recall that hands back impressions until the model asks `memory_get` for the full text
Developers who run the same project across many DeepSeek Harness sessions and keep re-deriving the same conclusions
Turn team conventions and procedures into rules the model actually sees every turn
A persistent rules block in the system prompt (`kind: rule`), a procedure index keyed by workspace and trigger words, and optional hook scripts that can reject a save, observe it, or rewrite / drop it before injection
Teams that want hard rules and SOPs enforced rather than re-explained in every session
Review, search, back up and audit everything the memory system did
A five-tab panel (service / records / memory tree / settings / audit) with a semantic-search switch, scope-level purpose and alias editing, and plain-JSON export that carries no vectors
Users who want to check by hand what was stored, what was recalled, and take the data with them
02How to install into DeepSeek Harness
Prerequisites
- Node.js 22.13 or newer — the README's Node.js badge declares ≥22.13
- A DSH profile to install into: desktop (`~/.dsh/profiles/desktop`) or web (`~/.dsh/profiles/web`)
Installation steps
- 01
$ dsh plugin --profile web add @restregular/dsh-memory
- 02
In the profile directory: `pnpm add @restregular/dsh-memory` — the package ships a `cordis.patch.yml` that registers the single entry.
- 03
Open 设置页 →「记忆系统」, or switch on the sidebar entry from 设置 → 侧栏入口.
- 04
For semantic recall, install the local llama.cpp backend from the panel's 服务 → 向量服务 tab; without it retrieval stays lexical.
Verify the integration
Not specified by the author
Rollback
- pnpm remove @restregular/dsh-memory
- Then delete the `dsh-memory` entry from the profile's patch.
- Uninstalling leaves the vector service and the memory files in `<storeDir>` / `<installRoot>`; clear them from the panel's 服务 → 维护 tab before you uninstall.
03DSH integration and capability boundaries
Native DSH host plugin: a single `cordis.patch.yml` entry in the profile provides the `memory` service, eight `memory_*` tools, the `agent/pre-step` injection chain and the `/memory/api` panel, while the client half registers the 「记忆系统」 settings page.
Eight memory tools for the model
content to save, query strings, scope names, record ids→impressions (8-char id / scope / kind / tags / short summary) from `memory_search`, full text and metadata from `memory_get`, plus scope directory, block and hook management
writes JSONL records under `storeDir`records what was returned in the session dedup stateDual-channel retrieval (lexical + local vectors)
a query string→memories ranked by relevance, with a relevance score when the vector channel is active
calls the local llama-server `/v1/embeddings` endpoint when one is configuredwrites a `.vec` sidecar next to each scope fileAutomatic recall and injection
the current user question, the first step of a session, tool failures and tagged behaviours→injected digests, an impression per memory, a resident rules block and a procedure index
inserts memory summaries into the system prompt and the conversation contextwrites the per-session dedup state under `<cwd>/.dsh/dsh-memory/sessions/`Per-session memory permissions with out-of-scope approval
the scope checkboxes, the read / write flags and the restriction level (absolute / normal / unrestricted)→tools only touch the authorised scopes; out-of-scope hits return a scope-only index and an approval card appears above that session's composer
writes `<storeDir>/visibility/<session>.json` and `<storeDir>/grants/<session>.json`blocks the session row with a pending-approval badge while the card waitsRule code (hook scripts)
a JS snippet with a `hookPhase` of `before-save`, `after-save` or `before-inject`→a rejected save with a reason handed back to the model, an observation note, or a rewritten / dropped memory before it reaches the context
stores code in `<storeDir>/hooks.json`, outside retrievalruns in a node:vm sandbox with no require / process / network and a 120ms timeout
04Who is it for? When not to use it?
Good for
- Developers who run the same project across many DeepSeek Harness sessions and keep re-deriving the same conclusions
- Teams that want hard rules and SOPs enforced rather than re-explained in every session
- Users who want to check by hand what was stored, what was recalled, and take the data with them
Not for
- Memory data stays on this machine as unencrypted JSONL, and exported files are plain text too — protecting the store directory is your job. The vector endpoint defaults to loopback.
05Compatibility, maintenance and safety notes
- Memory data stays on this machine as unencrypted JSONL, and exported files are plain text too — protecting the store directory is your job. The vector endpoint defaults to loopback.
- Audit entries live in memory only by default: capacity 2000 in total and 200 per session, keeping the digest / rules / delivery / error channels first — they are gone after a restart.
- Same-content dedup is still decided inside one process, so two processes writing exactly identical content at the same time can in theory leave two records. Cross-process write locking itself is per scope.
- The plugin only touches the network when an `endpoint` is set, `spawn` is configured, or the local backend was installed from the panel; otherwise it sends no requests and retrieval stays lexical.
MIT · README on branch main · no tagged release, last push 2026-11-12
06Frequently asked questions
Do I need a vector model for dsh-memory to work?
No. The README states that without a vector service retrieval degrades to lexical search (IDF plus Chinese bigram) and no network requests are made. Semantic recall needs a local llama-server backend, which the panel can install in one click with per-model `dims` and `pooling` already paired.
Where does dsh-memory put my data?
Under `storeDir`, which defaults to `.dsh-memory` relative to the launch directory — the README recommends configuring an absolute path instead. Records are plain JSONL with `.vec` sidecars, scope metadata lives in `scope-meta/`, and the panel settings go to `settings.json`.
How do I install it into a DSH profile?
Run `dsh plugin --profile web add @restregular/dsh-memory`, or add the package to the profile's dependencies manually. Either way the package's own `cordis.patch.yml` registers exactly one entry. The README explains that the host allows only one package name and one resolution source, which is why the panel and the server side live in a single package.
What happens if a new session still cannot see my memories?
Check that `inject.recall` / auto-recall is on, that the current scope is inside the recall range (tag-triggered recall only covers global, the current session and the current workspace), and that the tags really matched. The panel's 审计 tab shows candidates, filtering, budget and delivery status for each injection.
How do I remove it, and does that delete my memories?
Run `pnpm remove @restregular/dsh-memory` and delete the entry from the profile patch. Uninstalling does not remove the configured vector service or the memory files — they live in `<storeDir>` and `<installRoot>`, so clear them from the panel's 服务 → 维护 tab first, because the panel is gone once the plugin is uninstalled.
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
包自带 `cordis.patch.yml`,注册**唯一一条** entry:
面板分五个标签:**服务 / 记录 / 记忆树 / 设置 / 审计**。
This page is generated from the project’s public documentation, repository metadata and a structured parse of DSH Plugins; last verified on 2026-11-18. Found an error? Submit a correction.
Best DeepSeek Harness Plugins
Twelve plugins worth installing first — picked from the whole catalog, across every category.
