Introduction

DeepSeek Harness (DSH) adopts the “everything is a plugin” ecosystem philosophy. In agent development, memory modules often face two problems: either a complex JSON structure combined with a confidence machinery, or a lack of cross-session semantic integration. The dsh-evolving-memory plugin aims to solve this problem. It layers on top of the official compaction (in-session compression) and uses a low-cost model to further extract and consolidate compressed session summaries into human-readable Markdown files.

Plugin Overview

This plugin is named dsh-evolving-memory and is maintained by findshan. It provides file-based, self-evolving memory capabilities. The core workflow is: capture session content → dream consolidation → retrieval injection → evolution. Its design principle is to treat memory as editable documents rather than a closed database. The system only provides files, a retriever, and consolidation triggers; all judgments are left to the model.

Core Features

The plugin provides six core tools for reading, writing, and consolidating memory:
1. memory_search: Retrieve memory content.
2. memory_read: Expand memory details.
3. memory_catalog: Generate a memory catalog.
4. memory_save: Write to memory.
5. memory_correct: Learn through correction.
6. memory_dream: Perform cross-session consolidation.

Installation and Activation

Install and activate the plugin in the DSH environment:

dsh plugin --profile web add dsh-evolving-memory
dsh --profile web

Typical Usage

Memory files are saved by default in the $DSH_HOME/memory/ directory, with the following structure:

├── daily/2026-08-16.md   # 今日要点 + 会话纪要
├── user.md               # 用户画像（身份/偏好/目标/禁忌）—— 用户可编辑
├── agent.md              # 智能体自我认知与工作方式
├── memory.md             # 长期记忆（兜底落点）
├── dream.md              # 梦境整合日志
└── projects/<p>/project.md

When using it, invoke memory_search (search) and memory_read (expand) as needed for retrieval. Users can edit the user.md file. Agent writes are only suggestions, and the user has final decision-making authority. The dream consolidation gate defaults to either 24 hours or 5 sessions.

Notes

  • No schema, no migration, no confidence scores, no state machine: The system does not enforce structure, does not maintain migration scripts, does not output confidence scores, and does not use a state machine.
  • Human–agent co-governance: Agent writes are only suggestions, and the user has final decision-making authority.
  • Behavior when no API Key is available: Without an API key, memory read/write and retrieval continue as normal; only model-dependent steps are skipped.

Summary

dsh-evolving-memory simplifies the memory layer into a filesystem and uses official compaction events to drive extraction, achieving daily episodic layers and semantic memory consolidated through dream integration. It is suitable for developers who want full control over agent memory and prefer managing it as files.