@deepseek-ai/dsh-tool-memory

A persistent memory plugin for DeepSeek Harness (DSH) that enables agents to store and recall information across sessions, similar to the memory system in Hermes agent AI.

Features

• 🧠 Persistent Storage: Memories are saved to disk and survive across sessions
• 🔍 Flexible Search: Hybrid keyword+semantic (token Jaccard + substring) with mode: hybrid|keyword|semantic
• 🏷️ Categorization: Organize memories with optional categories (kebab-case, defaults to general)
• 📋 Seven Tools: memory_store, memory_search, memory_get, memory_list, memory_delete, memory_clear, memory_stats
• 🔄 Atomic Writes: write+rename with mkdir -p, corrupt-file recovery to *.corrupt.*
• ⚡ Performance: In-memory mtime cache (no re-read if unchanged), MAX_FACTS cap, timeoutMs:5000, isConcurrencySafe for reads
• ✅ Validation & Safety: kebab-case keys, length limits, memory_clear requires confirm:true, exec.signal abort handling
• 🧩 Modular & DSH-Native: lib/config|storage|validation|search/scoring|tools/*, peerDependencies to avoid dual-instance prepare bug

Installation

This package ships a dsh.bundle manifest, so it can be installed as a regular profile bundle.

As a DSH Plugin

1. Add the plugin to your DSH profile (this runs pnpm add in the profile directory, so a git URL works):

   dsh plugin --profile <your-profile> add github:disc0nct/dsh-memory-plugin

2. Register it as a bundle in the profile's package.json ($DSH_HOME/profiles/<your-profile>/package.json):

   {
     "dependencies": {
       "@deepseek-ai/dsh-tool-memory": "github:disc0nct/dsh-memory-plugin"
     },
     "dsh": {
       "profile": {
         "bundles": [
           "@deepseek-ai/dsh-base",
           "@deepseek-ai/dsh-web-app",
           "@deepseek-ai/dsh-tool-memory"
         ]
       }
     }
   }

3. Boot the profile:

   dsh --profile <your-profile>

Usage

Available Tools

Once installed, the following tools become available to your DSH agent:

memory_store

Save an important fact to long-term memory.

// Store a user preference
await ctx.tools.memory_store({
  key: "user-name",
  value: "Alice",
  category: "preferences"
});

// Store project information
await ctx.tools.memory_store({
  key: "project-language",
  value: "TypeScript",
  category: "project"
});

// Store a decision
await ctx.tools.memory_store({
  key: "api-decision",
  value: "Use REST API for simplicity",
  category: "decisions"
});

memory_search

Search for memories by keyword, category, or semantic paraphrase (hybrid keyword+token Jaccard ranking, dependency-free).

// Search all memories
const results = await ctx.tools.memory_search({
  query: "Alice"
});

// Search by category
const results = await ctx.tools.memory_search({
  category: "preferences"
});

// Combined search
const results = await ctx.tools.memory_search({
  query: "API",
  category: "decisions",
  limit: 5
});

// Semantic paraphrase: "fav color" matches "favorite-color"
const results = await ctx.tools.memory_search({
  query: "fav color",
  mode: "hybrid" // | "keyword" | "semantic" (default: "hybrid")
});

// Force exact substring only
const results = await ctx.tools.memory_search({
  query: "color",
  mode: "keyword"
});

memory_get

Fast exact lookup by key (vs memory_search scan).

const { found, fact } = await ctx.tools.memory_get({ key: "user-name" });
if (found) console.log(fact.value);

memory_list

List all stored memories (most recent first, optionally filtered).

// List all memories
const memories = await ctx.tools.memory_list();

// List memories by category
const memories = await ctx.tools.memory_list({
  category: "project"
});

memory_delete

Delete a specific memory by its key.

await ctx.tools.memory_delete({
  key: "user-name"
});

memory_clear

Clear ALL stored memories (requires explicit confirmation).

// cancelled without confirm
await ctx.tools.memory_clear(); // { cleared:false, count: N }

// confirmed
await ctx.tools.memory_clear({ confirm: true }); // { cleared:true, count: N }

memory_stats

Get health stats (count, per-category, oldest/newest, file size).

const stats = await ctx.tools.memory_stats();
console.log(stats.count, stats.categories); // {count: 12, categories:{project:5}}

Memory Storage Format

Memories are stored in ~/.dsh/memory.json with this structure:

{
  "facts": [
    {
      "id": "unique-identifier",
      "key": "user-name",
      "value": "Alice",
      "category": "preferences",
      "timestamp": "2024-01-15T10:30:00.000Z"
    }
  ]
}

Configuration

The memory file defaults to $DSH_HOME/memory.json (or ~/.dsh/memory.json when DSH_HOME is unset). Override it in the profile's patch layer ($DSH_HOME/profiles/<your-profile>/cordis.patch.yml):

- id: tool-memory
  config:
    memoryPath: /absolute/path/to/memory.json

Examples

Remembering User Information

// When user introduces themselves
if (userMessage.includes("my name is")) {
  const name = extractName(userMessage);
  await ctx.tools.memory_store({
    key: "user-name",
    value: name,
    category: "identity"
  });
}

// Later, when needing to address the user
const memory = await ctx.tools.memory_search({
  query: "name",
  category: "identity"
});
if (memory.results.length > 0) {
  await ctx.tools.memory_store({
    key: "greeting-used",
    value: `Hello ${memory.results[0].value}!`,
    category: "interaction"
  });
}

Project Context Tracking

// When starting work on a project
await ctx.tools.memory_store({
  key: "project-start",
  value: `Started work on ${projectName} at ${new Date().toISOString()}`,
  category: "project"
});

// When making a technical decision
await ctx.tools.memory_store({
  key: "tech-decision-db",
  value: "Selected PostgreSQL for reliability",
  category: "decisions"
});

// Later, when continuing work
const projectInfo = await ctx.tools.memory_list({
  category: "project"
});

How It Works

The plugin implements persistent memory by:

1. File Storage: Atomic writeFile(tmp)+rename to ~/.dsh/memory.json (no double-write), mkdir -p, max 5000 facts
2. Performance: In-memory mtime+size cache in lib/storage.js:8-78 — no re-read if file unchanged, clone on return, save updates cache
3. Efficient Lookups: Hybrid search via lib/search/scoring.js:12-137 (token Jaccard + substring boosts) then ISO timestamp desc; empty query → recency
4. Upsert Behavior: memory_store replaces existing memories with the same key and moves it to most-recent
5. Validation: key/category kebab-case (^[a-z0-9]+(-[a-z0-9]+)*$), value ≤10000 chars, category ≤32 chars
6. Recovery: On JSON SyntaxError, backs up to memory.json.corrupt.<ts> and returns empty store
7. Modular Layout: lib/config.js, lib/storage.js, lib/validation.js, lib/search/scoring.js, lib/tools/* (DSH apply re-exports)
8. DSH Idioms: Config via @deepseek-ai/schemastery, defineTool with timeoutMs:5000, isConcurrencySafe for reads, kind hints, exec.signal abort, peerDependencies

Requirements

• DeepSeek Harness (DSH) v0.1.0-rc.8 or later
• Node.js v18.0.0 or later (uses crypto.randomUUID, fs/promises.rename/stat)
• Peer dependencies:
  • @deepseek-ai/cordis: ^4.0.1
  • @deepseek-ai/dsh-tools: ^0.1.0-rc.8
  • @deepseek-ai/schemastery: ^3.18.1

License

MIT License - feel free to use, modify, and distribute this plugin.

Development

To contribute to this plugin:

1. Fork the repository
2. Create a feature branch
3. Make your changes
4. Ensure all tests pass (if applicable)
5. Submit a pull request

Credits

Inspired by the memory systems in agents like Hermes AI, this plugin brings similar long-term memory capabilities to the DeepSeek Harness ecosystem.

Built with ❤️ for the DSH community