In the development practice of DeepSeek Harness (DSH), agents often face context discontinuity. After a session ends, previously accumulated experience cannot be used by subsequent conversations, leading to duplicated work. dsh-remember aims to solve this problem. It does not rely on external vector databases or third-party services; instead, it stores memory as local Markdown files, ensuring continuity and controllability.

This is a plugin for DeepSeek Harness (dsh) that provides cross-session long-term memory. Maintained by user Arborsm. It scans sessions in the background, extracts and integrates useful experience, and automatically injects summaries into later conversations, allowing the model to retrieve original details when needed.

Core Features

  • Automatic Extraction: Scans sessions in the background and extracts useful information.
  • Automatic Integration: Incrementally merges new experience and removes stale entries that are no longer referenced.
  • Automatic Recall: Injects a memory summary at the start of each conversation (with a configurable token limit), and the model retrieves the original text as needed.
  • Readable and Controllable: Memory files are plain Markdown and can be manually viewed, searched, and managed.
  • Translatable: Supports one-click translation of memory files and retains the incremental cache.
  • Portable: Supports gzip packaging for import/export and supports content-hash-based de-duplication.

Installation and Enablement

After installing the plugin, restart dsh web to activate it.

dsh plugin --profile <name> add dsh-remember

Typical Usage

Once the plugin is running, memory data is stored under ~/.dsh-memory/. This directory contains a git workspace, and the files in it can be directly viewed, searched, and managed using a text editor or git commands.

At the start of each new conversation, the model automatically receives a summary based on prior memory and retrieves the original files when details are needed. The whole process does not require configuring a vector database or external services.

Use Cases and Notes

Suitable for developers who need to accumulate long-term experience and do not want to rely on cloud vector databases or third-party services.

The plugin runs with the permissions of the current DSH process, so local file read/write security should be considered. The license is MIT, permitting free modification and redistribution.

Summary

dsh-remember uses the file system to implement persistent storage and version control of memory, allowing agent developers to manage their memory assets locally. For more configuration details and source code, visit its GitHub repository: https://github.com/Arborsm/dsh-remember.