creativedswork/dsh-expmem

DSH ExpMem

English | 简体中文

Experience Memory for DeepSeek Harness.

Introduction

DSH ExpMem is a community plugin for long-term personal memory in DeepSeek Harness. It keeps user habits, reusable task experience, and engineering insights in inspectable local files. DSH's existing Session history remains the verbatim Recall layer.

The design combines two research lines. MemGPT supplies the memory infrastructure: tiered storage, memory pressure, retrieval, and explicit memory operations. Generative Agents supplies the cognitive process: assign importance, retrieve by current relevance, synthesize reflections, and use those memories when planning the next action.

DSH ExpMem is not an official DeepSeek project.

Architecture

flowchart TB
  subgraph Runtime["DeepSeek Harness runtime"]
    Agent["DSH Agent"]
    Meter["Token Meter"]
    Sessions["Session Persistence<br/>JSONL"]
  end

  subgraph Memory["MemGPT-style memory infrastructure"]
    Recall["Recall<br/>verbatim session history"]
    Archive["ExpMem Archive<br/>local JSON records"]
  end

  subgraph Cognition["Generative Agents-style cognitive loop"]
    Retrieve["Retrieve<br/>relevance + recency + importance"]
    Reflect["Reflection Run<br/>pending → prepared → completed"]
    Plan["Plan and act<br/>inside the DSH agent loop"]
  end

  Sources["Claude Code / Codex<br/>Markdown memory"] -->|"idempotent import"| Archive
  Sessions --> Recall
  Meter -->|"70% context pressure"| Agent
  Recall --> Retrieve
  Archive --> Retrieve
  Retrieve --> Agent
  Agent -->|"promote durable experience"| Archive
  Agent --> Reflect
  Reflect --> Retrieve
  Retrieve --> Reflect
  Reflect --> Archive
  Agent --> Plan

DSH owns Recall, context compaction, and the Agent Loop. ExpMem owns the distilled Archive, retrieval ranking, reflection provenance, and the trustworthy memory lifecycle. It does not copy Session events.

Context Engineering

The design maps to four common Context Engineering operations: