DEV Community

Cover image for Soul v5.0: Your AI Agent Now Remembers People, Hardware, and Projects Across Sessions
N2 Dev
N2 Dev

Posted on

Soul v5.0: Your AI Agent Now Remembers People, Hardware, and Projects Across Sessions

#ai #mcp #productivity #opensource

Soul v5.0: Your AI Agent Now Remembers People, Hardware, and Projects

Every AI coding session starts with digital amnesia. Your agent doesn't know what it did yesterday. It doesn't remember your GPU is an RTX 5060 Ti. It doesn't know your backend runs on a Mac Mini. You re-explain everything. Every. Single. Time.

Soul v5.0 fixes this with three new memory systems.

🧠 What's New in v5.0?

1. Entity Memory

Soul now auto-tracks structured entities — people, hardware, projects, concepts — across sessions.

n2_boot output:
Entities(5): n2-browser[project]: v0.7.2, Electron + TypeScript
| master[person]: RTX 5060 Ti, 5600X, Windows
| mac-mini[hardware]: backend-server, Ollama, SearXNG
Enter fullscreen mode Exit fullscreen mode

No more re-explaining your setup. Your agent just knows.

2. Core Memory

Agent-specific facts that are always injected at boot — identity, rules, current focus. Think of it as the agent's "self-awareness layer."

Core[rose]: identity: Team lead, calm but caring
| current_focus: Soul v5.0 memory upgrade
| rules: PowerShell uses ; not &&
Enter fullscreen mode Exit fullscreen mode

Each agent gets its own core memory. Rose knows she's the team lead. Jenny knows she handles design. They never confuse their roles.

3. Autonomous Extraction

At session end, entities and insights are auto-saved without explicit calls:

n2_work_end({
  agent: "rose",
  project: "my-app",
  title: "Implemented auth",
  summary: "Added JWT auth with refresh tokens",
  // 🆕 Auto-extracted
  entities: [
    { type: "concept", name: "JWT", attributes: { lib: "jose" } }
  ],
  insights: [
    "Refresh token rotation prevents replay attacks",
    "jose library is 3x faster than jsonwebtoken"
  ]
});
Enter fullscreen mode Exit fullscreen mode

Next session, those insights are searchable. That knowledge compounds.

⚡ 2-Minute Setup 

npm install n2-soul
# or
git clone https://github.com/choihyunsus/soul.git
Enter fullscreen mode Exit fullscreen mode

Add to your MCP config:

{
  "mcpServers": {
    "soul": {
      "command": "node",
      "args": ["/path/to/soul/index.js"]
    }
  }
}
Enter fullscreen mode Exit fullscreen mode

Tell your agent:

- Start of session: call n2_boot(agent_name, project_name)
- End of session: call n2_work_end with summary, TODO, entities, insights
Enter fullscreen mode Exit fullscreen mode

That's it. Two commands. Your agent now has persistent memory.

🔄 How the Cycle Works 

┌─────────────────────────────────────────────┐
│  BOOT                                       │
│  n2_boot(agent, project)                    │
│  → Entity Memory loaded (people, hardware)  │
│  → Core Memory injected (identity, rules)   │
│  → Handoff from previous session            │
│  → KV-Cache context restored                │
└──────────────────┬──────────────────────────┘
                   ↓
┌─────────────────────────────────────────────┐
│  WORK                                       │
│  n2_entity_upsert  → track new entities     │
│  n2_core_write     → save agent facts       │
│  n2_brain_read/write → shared memory        │
│  n2_work_claim     → prevent file conflicts │
└──────────────────┬──────────────────────────┘
                   ↓
┌─────────────────────────────────────────────┐
│  END                                        │
│  n2_work_end(agent, project, ...)           │
│  → Immutable ledger saved                   │
│  → Entities auto-saved                      │
│  → Insights archived                        │
│  → Handoff ready for next agent             │
└─────────────────────────────────────────────┘
Enter fullscreen mode Exit fullscreen mode

📊 New Tools in v5.0

Tool What it does
n2_entity_upsert Add/update entities with auto-merge
n2_entity_search Search by keyword or type
n2_core_read Read agent's core memory
n2_core_write Write persistent agent facts

These join the existing 14 tools (boot, work tracking, brain, KV-Cache, context search, backup/restore).

🏗️ Built for Real Multi-Agent Workflows

Soul isn't an experiment. It runs in production inside N2 Browser — an AI-native browser with 4 concurrent agents:

  • Rose (Antigravity) — team lead, architecture
  • Jenny (Gemini Flash Lite) — design, quick tasks
  • Jisoo (Gemini Flash) — coding, debugging
  • Lisa (Qwen 8B) — research, local inference

Each agent has its own Core Memory. Entity Memory is shared. Handoffs are seamless. No agent ever asks "who are you?" twice.

🔗 Links

What's Next?

  • Semantic search with Ollama embeddings (config-ready, opt-in)
  • Auto-pruning for entities that haven't been mentioned recently
  • Cross-project entity linking — same person across different projects

Soul v5.0 was verified, documented, and published by Rose — an AI agent running inside N2 Browser. The code was written, tested (16/16 PASS), and this post drafted in a single session. Agents building tools for agents. How meta is that?

Star us ⭐ if your AI deserves a memory.

Top comments (0)