pacoyi/dsh-memory-lite

dsh-memory-lite is a native lightweight cross-session memory plugin built exclusively for DeepSeek Harness (DSH). It adds a memory tool with four core operations (save, recall, list, forget) and a dedicated “记忆库” (Memory Library) card in the DSH settings interface that includes an import wizard, export panel, scope isolation, and approval-gated writes. To install the plugin, you need to have the DSH CLI installed first, then run dsh plugin --profile [profile-name] add github:pacoyi/dsh-memory-lite to add it to your desired DSH profile, then restart the profile to activate it. It has zero third-party dependencies, only requiring two peer dependencies from the official DSH codebase: @deepseek-ai/cordis and @deepseek-ai/dsh-tools.

By default, every new DSH session starts with a blank slate, and the model cannot access any long-term facts stored in past session transcripts. This plugin fills that gap by letting the agent save and recall durable facts like user preferences, project decisions, and key constraints that need to persist across restarts, upgrades, and profile changes. All memory data is stored locally in a human-readable JSON file at ~/.dsh/memory-lite.json, which you can open, edit, back up, or delete manually when the DSH host is stopped. While the plugin itself makes no network requests and has no telemetry, any recalled memory is sent to your configured model provider, so you should never store sensitive information like credentials, API keys, or private source code in it.

The plugin includes multiple security and durability features, such as scope isolation that prevents one project from reading or modifying another project’s memory entries, atomic writes that avoid data corruption, and a full audit trail for all memory changes. All save and forget operations require explicit user approval before execution, and deleted entries are moved to a trash bin that lets you restore them later until you confirm permanent deletion. It enforces clear budget limits: notes are capped at 2000 characters, there is a maximum of 1000 entries total, and the full storage file is capped at 512 KB. The project is open source under the permissive MIT license, and all design choices are documented for maintenance.