Introduction

Building agents using frameworks like DSH (DeepSeek Harness), where “everything is a plugin”, requires building long-term memory from scratch: deciding where to store it, how to compress it, and how to read it next time—every step involves manual work. A diary is a typical scenario: you need to archive the original text while ensuring the model can grasp the key points when reviewing it the next day, without starting with a cold statistical report.

Below is an introduction to dsh-diary-detective (Cyber Diary Detective), which packages diary read/write operations, memory compression, and gentle feedback into a single plugin. The project is maintained by SageChenAI and is licensed under MIT.

What is it

One-sentence positioning: A warm mirror that reads your diary, reflects blind spots, and provides gentle feedback.

It operates in two layers:

  1. Mechanical Layer: Stores spoken diaries as files, reads back the original text, performs basic statistics, and maintains a compressed memory.
  2. Understanding Layer: Driven by the “Warm Observer” guidelines in SKILL.md, it instructs the model on how to think, analyze, and speak.

It reflects both blind spots (repeatedly made mistakes and their root causes) and areas for growth. Its identity boundaries are clearly defined: it is not a therapist, a coach, nor a productivity tracker. Analysis is provided in the form of hypotheses, which users can confirm, deny, or correct.

Core Features

The plugin provides five tools:

  • diary_write: Saves spoken diaries as YYYY-MM-DD.md, without enforcing a specific format.
  • diary_read: Reads the original diary entries; the original text is the sole source of truth.
  • diary_index: Mechanical statistics, including day-of-week distribution, tag frequency, Chinese bigram frequency, consecutive days, and gaps—note that these are only rough proxies; semantic understanding is the model’s responsibility.
  • diary_memory / diary_memory_update: Read and rewrite compressed memory.
  • diary_feedback: An entry-level package for a single call, returning today’s entry, memory, and statistics together.

Beyond the tools, there is a skill: SKILL.md, the “Warm Observer” guidelines. It constrains the manner of expression—discussing only one discovery at a time and citing original text evidence; stating plainly when there is nothing worth saying instead of fabricating patterns for the sake of it; listening first, then expressing.

Data is stored relative to the workspace by default: diaries in <workspace>/diary/ and memory in <workspace>/diary-memory.md. All tools support dir / path parameters, allowing you to specify where to store them.

Installation and Activation

First, install the tool (bundle):

dsh plugin --profile <name> add dsh-diary-detective

Replace <name> with your profile name. You can also install from GitHub, replacing the placeholders in the command yourself:

dsh plugin --profile <name> add github:<your-github-username>/dsh-diary-detective

Next is the second step: manually install the skill. Copy the SKILL.md from the repository to the skills directory:

# User-level
cp SKILL.md ~/.dsh/skills/diary-detective/SKILL.md
# Or project-level
cp SKILL.md <project>/.dsh/skills/diary-detective/SKILL.md

After the steps above, both tools and skills are in place. Current version is 0.1.0, depends on @deepseek-ai/dsh-tools ^0.1.0. Skills currently require manual copying; Roadmap v1.1 plans to support automatic skill registration via bundles and plugin configuration paths.

Typical Usage

Daily usage is triggered by natural language:

  • “Today’s review” / “daily review”
  • “Write diary: today…” / “journal: today I…”
  • “Have I been… lately?” / “have I been… lately?”

Taking “Today’s review” as an example, the model will first load the skill, then call tools in sequence: read original text → read memory → mechanical statistics → analyze → express → update memory. It collects facts first, then performs understanding and expression, and finally writes new conclusions back to memory for the next use.

Privacy

  • All diary content is stored locally, with the location specified by you via dir / path
  • Original text may be sent to the model for analysis, but this project will not share it or use it for training

Use Cases and Notes

Suitable for two types of people:

  • DSH users who want to add long-term diary memory to their agents; the toolchain is ready-made, so no need to write storage and statistics yourself.
  • People sensitive to the tone of analysis. These guidelines emphasize not labeling, not fabricating patterns, and not digging up old scores; when the content exceeds the capability range, it suggests seeking a real person.

Two notes:

  • The plugin runs with the permissions of the current dsh process and can read/write paths it gains access to. It is recommended to review the source code and license before installing (this project is MIT).
  • The project is still in early stages (0.1.0). v1.1 plans to add diary_report (weekly/monthly review) and plugin configuration paths, while v1.2 plans a general prompt version (pasted into any AI) and a browser panel with consecutive day statistics. Please pay attention when making selections.

Summary

dsh-diary-detective packages “saving diaries, reading diaries, remembering conclusions, and expressing gently” into a single plugin: tools handle facts, SKILL.md handles boundaries. If you are supplementing diary and memory capabilities for your DSH agent, you can start with this.

  • Community Plugin Directory: https://www.skillhub.cn/plugins/SageChenAI/dsh-diary-detective
  • GitHub Repository: https://github.com/SageChenAI/dsh-diary-detective

(Note: skillhub.cn is a community-maintained plugin directory and has no official affiliation with DeepSeek / Hypothesis.)