@deepseek-ai/dsh-plugin-memory

English | 中文

Persistent five-layer memory for DeepSeek Harness: profile, project context, daily log, and recallable topics — so the agent remembers you across sessions, not just within one.







Two cordis seams — agent/pre-step injection + ctx.tools.register (six tools)

A persistent five-layer memory system for DeepSeek Harness (DSH): a user profile (L1), a per-project semantic index with topic files (L2), and append-only per-day logs (L3) under ~/.dsh/memory/ and <cwd>/.dsh/memory/. It injects relevant memories into every request and auto-extracts durable facts from finished sessions via the LLM. Integrates as a DSH plugin on two cordis seams — agent/pre-step for injection, ctx.tools.register for six memory_* agent tools. Without an llm route it still works: entry injection, keyword relevance, and profile rotation remain; only LLM ranking and auto-extraction are disabled.

✨ Features

• 🧠 Five-layer model: L0 user-owned identity (~/.dsh/AGENTS.md, not managed by the plugin) → L1 profile → L2 project index + topics → L3 per-day append-only log → L4 skills (existing). Each layer has its own write path, truncation budget, and injection rule.
• 📇 Index + topic split (L2): MEMORY.md is always an index of one-line pointers (≤150 chars each); details live in <topic>.md. Keeps single files small, searchable, and truncatable.
• ✂️ Truncation budget: the booted index is hard-clamped to 200 lines / 40,000 chars, so cold-start context stays cheap.
• 🎯 Relevance injection: on each step, the latest user query selects relevant topic files (LLM ranking when llm is configured, keyword scoring otherwise) and appends them as a <system-reminder data-role="memory"> block; files already surfaced in this session are de-duplicated. The two channels are labeled memory-entry (once per session) and memory-relevance (per step) in the GUI context rows.
• 🤖 LLM auto-extraction: when a session goes idle, a debounced (60 s) best-effort pass scans the recent 40 events, asks the LLM for new topic files and index lines, and writes them. Never overwrites existing memories; degrades silently if the model is unavailable.
• 🔄 Profile rotation (L1): memory_profile merges new facts into four fixed sections (工作背景 / 个人背景 / 当前关注 / 近期动态) and rotates the version, keeping the previous copy in profile.md.bak.
• 🔒 Read-back data, not instructions: memory is written with fs/promises directly to the memory roots — intended persistence, not self-modification — and paths are confined to the store root. Memory files are context the agent reads back, never permission grants.
• 🧩 Pure harness plugin: no HTTP API or GUI panel — injection and tools only. DSH serves a single user, so paths carry no <uid> layer.
• 🛠️ Six agent tools registered via ctx.tools.register (defineTool from @deepseek-ai/dsh-tools):

Quick Start

Prerequisites

• A DeepSeek Harness (DSH) installation with a plugin-capable profile (e.g. web).
• No LLM route required — the plugin falls back to keyword-only relevance.

Install

dsh plugin --profile web add github:NattoCB/dsh-plugin-memory

Run

Restart dsh web. On first use the plugin bootstraps both memory roots:

~/.dsh/memory/
  MEMORY.md        # global index (≤200 lines / 40K chars)
  profile.md       # L1 profile (Version N)
  profile.md.bak   # previous profile version
  <topic>.md       # global topic files
<cwd>/.dsh/memory/
  MEMORY.md        # project index
  YYYY-MM-DD.md    # daily memory (append-only)
  <topic>.md       # project topic files

Tell the agent something worth remembering, or let idle auto-extraction pick it up — then check the memory roots a session later.

Configuration

Deploy the plugin via a DSH bundle entry (see cordis.patch.yml and package.json exports):

Example entry:

- id: memory
  name: '@deepseek-ai/dsh-plugin-memory'
  config:
    enableEntryInjection: true
    enableRelevance: true
    enableExtraction: true
    maxRelevant: 5
    relevanceTopK: 8
    relevanceBudgetChars: 2000
    extractionDebounceMs: 60000
    extractionLookback: 40
    llm:
      provider: deepseek   # example: fill in your route
      model: deepseek-chat
      maxTokens: 1024

License

MIT — see LICENSE.