dsh memory plugin: install & recall in DeepSeek Harness

A dsh memory plugin is the DSH plugin that gives DeepSeek Harness cross-session memory: it solves a concrete problem — open a new session and the model no longer remembers the project background, naming habits and output preferences you set in the last one (source). This guide covers three things: what it solves and how it differs from built-in session memory; how to find, install and enable one in the DSH Plugin Hub; and how memory is written, recalled and kept private. To compare several memory plugins side by side, see choosing a DeepSeek Harness memory plugin.

Overview

In one line: a dsh memory plugin moves the things worth remembering from in-process state onto disk, so a new session opens with the previous session's context. It does not raise the quality of a single turn; it changes persistence and recall — which information deserves to stay, where it lives, and how the next session pulls it back. What follows first covers the problem it solves, then the install and enable steps in the Hub, then the read/write and privacy boundary.

What a dsh memory plugin is: cross-session memory vs built-in session history

The core value of a dsh memory plugin is cross-session long-term memory: without it, DeepSeek Harness session state lives only in the process in some modes, so a restart or a new session loses it (source). Built-in session history records "what we talked about this time"; a memory plugin records "what should be remembered from now on." The difference looks like this:

Such plugins generally do four jobs: automatic extraction (pull durable facts from a conversation), automatic injection (put relevant memory into a new session's context), periodic consolidation (dedupe, merge, compress) and local persistence (write to a local SQLite database or files). A memory plugin is therefore a reusable long-term thread for the model, not a replay of the whole chat.

How to install and enable a dsh memory plugin in the DSH Plugin Hub

Installing a dsh memory plugin always goes through the DSH Plugin Hub: search the market for the plugin, click install and confirm, then restart so the plugin loads. DSH Plugin Hub is the official market built into Settings → Plugin Market, where search, install, update and uninstall all happen. Using a real memory plugin, dsh-memory, the steps are:

1. Check prerequisites: DeepSeek Harness is running and the plugin market is installed; if not, install it first with how to install a DSH plugin. Expected: a "Plugin Market" entry appears on the settings page.
2. Open the market and search: open the Web UI → Settings → Plugin Market, and type memory or the plugin name into the search box. Expected: the list filters down to memory plugins.
3. Click install: on the target plugin card, click Install. Expected: a confirmation dialog opens.
4. Verify the dialog: the dialog shows the plugin name, source repository and the dsh plugin --profile web add <package> command it will run; confirm once it checks out. Expected: the download and install begin in the background with live progress.
5. Restart and enable: stop and start dsh web so the plugin reloads. Expected: the card turns to "Installed" and the plugin's capability takes effect.
6. Verify persistence: open a new session and restate a preference from the previous one. Expected: the model answers with that old context; if not, return to the market and confirm the plugin is enabled and memory has started writing locally.

The command line is the equivalent second path: dsh plugin --profile web add <package> (source). Both paths end the same way; the only difference is a button versus a command.

How a dsh memory plugin stores and recalls memory — and the privacy limits

A memory plugin's data lands locally by default (SQLite or files), and writing and recall follow two mechanisms: distill during the conversation, persist to disk, inject by relevance in a new session. Break it into three points:

1. Where it writes — most implementations store memory in local files or a SQLite database, usually under the plugin's own data directory; that means backup, migration and cleanup all have to be handled separately, and uninstalling the plugin does not erase memory already written.
2. How it is recalled — when a new session starts, the plugin retrieves historical key points against the current question and injects the most relevant fragments into context; some use keyword retrieval (such as BM25), some semantic vectors, some a knowledge graph, and the retrieval method decides which memories surface.
3. What to watch for privacy — memory persists what was said, which means conversation content stays on disk long term; with keys, customer data or private preferences, first confirm the storage location, whether there is approval gating, and whether a single memory can be deleted by hand.

One line to remember: the cost of "remembering" is "writing it down," and what is written down should be handled as sensitive data.

Rather than tracking a memory plugin's install, update and uninstall through scattered commands, use the DSH Plugin Hub Installed list to manage them in one place: what is installed, whether a newer version exists, and whether to remove it, all on one screen.

Notes and limits

Treat a dsh memory plugin as a long-term notebook rather than an all-purpose memory, and it is easier to live with. A few boundaries:

1. Installing is not the finish line: recall depends on retrieval quality, so a vague question may pull back nothing — the more specific the wording, the better it lands.
2. Do not run several of the same kind: memory plugins overlap heavily, and installing more than one causes duplicate injection and interference; keep a single primary memory plugin.
3. Uninstalling does not delete data: removing the plugin only clears its dependency; memory files already written need separate cleanup.
4. Be careful with sensitive content: with private data or credentials, check the storage location and the deletion path before letting the plugin record.
5. Check compatibility: prefer entries marked as verified in the market, and trial anything unconfirmed on a small scale first.

Sources: DeepSeek Harness 官方文档 - 会话子系统, dshplugin/dsh-plugin-hub