dsh-memento: cross-session memory for DeepSeek Harness
dsh-memento
A DeepSeek Harness plugin that keeps notes between sessions in local SQLite, each write approved and logged in an audit trail
- 144 ★ on the repository
- Last commit
- Apache-2.0
Extended access
Why this level
- The plugin writes notes to disk and mixes them into the system prompt of later sessions: a wrong entry shapes the agent's behaviour
- In `writePolicy: auto` mode writes go in without confirmation
Install
Manual install
dsh plugin --profile web add dsh-mementoThe npm version. For the latest main: dsh plugin --profile web add github:PerryLink/dsh-memento.
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The plugin writes notes to disk and mixes them into the system prompt of later sessions: a wrong entry shapes the agent's behaviour
- In `writePolicy: auto` mode writes go in without confirmation
This is third-party code. Review the repository files before installing.
What it does
dsh-memento adds a `ctx.memory` service and a local SQLite store (`node:sqlite`, WAL mode, `0600` permissions) to DeepSeek Harness. The model saves notes with the `memory` tool and receives them as a frozen snapshot in the system prompt at the start of the next session. Notes live on two tracks, `user` and `agent`, and on two layers, user-global and workspace. Each pair has a hard character budget: 2000 for `user` and 4000 for `agent`. A write longer than the budget is rejected before it is stored. Every change goes through an approval gate. The `writePolicy` setting takes `ask`, `auto` or `off`, and the audit trail consists of a pair of approval events, an audit table and snapshot rows. Search is a case-insensitive substring match, with vector search as an opt-in. The `/memory` command shows budgets and the list of notes, and the `dsh-memento-mcp` entry point exposes the memory read-only over stdio.
Who it is for. DeepSeek Harness users who want a memory they can audit: what the agent remembered is visible in a list, and nothing reaches the store without approval.
Good fit when
- The agent has to remember project agreements so the next session does not ask again
- Each memory write must be approved by hand and later matched against the log
- Memory must not grow without limit: each track and layer has a character cap
- Notes are needed separately for the user and for a specific workspace
- Another client needs read-only access to the memory over MCP
Not a fit when
- You work in Claude Code, Cursor or Codex: the plugin targets DeepSeek Harness only
- You need a store for large bodies of knowledge: a budget of a few thousand characters fits short notes
- You need shared cloud memory for a team: the data sits in a local file
- The agent should write to memory without prompts by default: `writePolicy` has to be switched to `auto`
Example request
Remember that tests in this project run with make test, and show what is stored in the workspace memory.Limitations
The plugin works with DeepSeek Harness 0.1.2 up to 0.3.0 and needs Node.js 22.19 or 24.9 and newer. It is a community plugin with no connection to DeepSeek. Default search matches substrings, so a reworded note may not be found.
Removing the plugin leaves the notes database and session logs on disk, and you delete them by hand. `writePolicy: auto` turns off prompts before writing, and the audit log then becomes the only check. Installing from the git branch pulls the latest main and does not pin a version.
How to disable. Remove the plugin with `dsh plugin --profile web remove dsh-memento`. The notes database and session logs stay; to clean up, delete `memory.db` in `$DSH_HOME/dsh-memento`.
FAQ
How do I install dsh-memento?
Run `dsh plugin --profile <name> add dsh-memento` for the npm version. For the latest main, use `dsh plugin --profile web add github:PerryLink/dsh-memento`. No build step is needed: `index.mjs`, `lib/` and `client/` are the release.
What is DeepSeek Harness?
It is an agent runtime with a `dsh` command and profiles, and plugins install into it. dsh-memento is built for it and does not run in other agents such as Claude Code.
Where are the notes stored and can I read them?
By default in `$DSH_HOME/dsh-memento/memory.db` with `0600` permissions. It holds tables for entries, audit, proposals and metadata. You can view the contents with `/memory list` in a session.
What happens when the budget is full?
The plugin rejects a write that does not fit the track and layer limit. The defaults are 2000 characters for `user` and 4000 for `agent`, and you change them in the `budgets` block of the config.
How does it differ from other DSH memory plugins?
The author describes it as local SQLite with an approval gate and an audit chain. The neighbours named in the README work differently: dsh-mnemon connects hosted stores, and the closest, dsh-plugin-memory, keeps no provenance log.
Discussion
No questions or reviews yet.
Related
PluginsLong-term memory for DeepSeek Harness with an MCP server: the same memories reachable from Claude Code, Cursor and Codex
A DeepSeek Harness plugin: recent talks stay detailed, older ones compress, and key facts live as events with sources.
Local memory for agents with a review queue: SQLite, an audit log and a token ledger, one vault for Claude Code, Codex and Cursor
A shared local SQLite memory for Claude Code, Codex and Hermes: every fact has a source and can be corrected or deleted
A DeepSeek Harness plugin for a standing team of agents with memory, projects, task owners and verification by a second member
Shared semantic memory for agents: conversations are saved as markdown and recalled by search across Claude Code, Codex and other agents
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| dsh-memento: cross-session memory for DeepSeek Harness (this card) | no | 144 | ||
| agent-memory | no | 4,315 | a separate memory CLI layer without a DSH approval gate | |
| Basic Memory | no | not required | 4,566 | Markdown notes over MCP for any MCP client |
README in short
The README covers three install routes, a minimal config and a check with `/memory budgets` and `/memory list`. Separate documents describe memory protocol v1, its JSON schema and an adapters guide. Third-party stores connect through adapters.