Role Persona

Role-based persona system for AI agents — memory, knowledge, embedding. CLI · HTTP Daemon · MCP Server · Pi Extension · Web Dashboard

---

Overview

Role Persona is a persistent memory and knowledge system for AI coding agents. It provides:

• Role Management — independent personas with isolated memory/knowledge
• Memory System — learnings, preferences, auto-extraction, consolidation, vector search
• Knowledge Base — multi-source, tag-searchable, version-controlled entries
• Vector Memory — LanceDB-backed semantic search with hybrid keyword+vector fusion
• Multiple Interfaces — CLI, HTTP daemon, MCP server, Pi extension, Web dashboard

Architecture

┌─────────────────────────────────────────────────────────────────┐
│                         Consumers                                │
│  ┌──────────┐  ┌──────────┐  ┌──────────┐  ┌──────────────────┐│
│  │   CLI    │  │  Daemon  │  │ MCP Srv  │  │   Pi Plugin      ││
│  │          │  │ (HTTP)   │  │ (RPC)    │  │ (direct service) ││
│  └────┬─────┘  └────┬─────┘  └────┬─────┘  └───────┬──────────┘│
│       │              │              │                │           │
│  ┌────┴──────────────┴──────────────┴────────────────┴────────┐ │
│  │              ServiceManager (CWD multiplexing)              │ │
│  │   CWD → Service instance    Role name → Service instance   │ │
│  │   30min idle timeout        Auto cleanup                    │ │
│  └────┬───────────────────────────────────────────────────────┘ │
│       │                                                         │
│  ┌────┴──────────────────────────────────────────────────────┐ │
│  │                   Service Layer                            │ │
│  │  RoleService · MemoryService · KnowledgeService            │ │
│  │  EmbeddingService · SystemPromptBuilder                    │ │
│  └────┬──────────────┬──────────────┬───────────────┬────────┘ │
│       │              │              │               │           │
│  ┌────┴──────┐ ┌─────┴─────┐ ┌─────┴──────┐ ┌─────┴────────┐ │
│  │  Core     │ │ Embedding │ │ Vector DB  │ │   Config     │ │
│  │ memory-md │ │ providers │ │  (LanceDB) │ │  (JSONC)     │ │
│  │ memory-llm│ │ OpenAI    │ │            │ │              │ │
│  │ memory-tag│ │ Local     │ │            │ │              │ │
│  └───────────┘ │ MiniLM    │ └────────────┘ └──────────────┘ │
│                └───────────┘                                  │
└─────────────────────────────────────────────────────────────────┘

Multiplexing

The daemon supports CWD-based and role-based multiplexing:

CLI mode:    POST /api/memory/list { "cwd": "/project" }  → resolves role from CWD
Web mode:    POST /api/memory/list { "role": "zero" }     → directly activates role
Pi mode:     Direct service import (no HTTP)

Each CWD/role gets an independent service instance with 30-minute idle timeout.

Quick Start

Installation

git clone https://github.com/Dwsy/role-persona.git
cd role-persona
bun install
bun run build

CLI Usage

# Initialize (auto-creates role from CWD)
role-persona init

# Role management
role-persona role list
role-persona role create my-role
role-persona role info

# Memory
role-persona memory list
role-persona memory search "query"
role-persona memory add-learning "Always use TypeScript strict mode"
role-persona memory add-preference "Use pnpm" --category Tools
role-persona memory consolidate
role-persona memory repair
role-persona memory tidy

# Knowledge
role-persona knowledge list
role-persona knowledge search "design patterns"
role-persona knowledge write --title "ADR-001" --content "..." --category architecture

# Vector memory
role-persona embedding stats
role-persona embedding rebuild

# Scenario memory (L2)
role-persona memory scenario-write --title "Code review output" --guidance "Give the conclusion first, then group findings by Critical/Medium/Low." --triggers "code review,review feedback"
role-persona memory scenario-search "review this PR"
role-persona memory scenario-list
role-persona memory scenario-read <scenario-id>

# System prompt
role-persona prompt --base "You are a helpful assistant."

HTTP Daemon

# Start (persists in memory, warm service)
role-persona daemon start

# Background mode
role-persona daemon start --background

# Custom port
role-persona daemon start --port 8080

# Status / Stop
role-persona daemon status
role-persona daemon stop

Web Dashboard

# Build
cd web && bun install && bun run build

# Start
role-persona --web                    # unified daemon + web on port 3939
role-persona --web --port 8080        # custom unified port

Features:

• Dashboard — role banner, memory breakdown, knowledge sources, vector stats, recent activity
• Memory — Explorer tree navigation, table view with edit/delete, Markdown rendering, regex search, keyboard shortcuts (/ search, j/k nav, c copy)
• Knowledge — list, search, click to view Markdown, write new entries
• Roles — click to view role detail modal, switch roles
• Settings — daemon status, vector stats, JSONC config editor (Form mode with schema-driven inputs + raw JSONC mode)
• i18n — English/Chinese
• Dark/Light — theme toggle
• Role selector — header dropdown, auto-refresh all pages on switch
• URL state — tab/path/search persisted in URL params (bookmarkable)
• Model selector — dropdown populated from models.json

Pi Extension

// extensions/role-persona/index.ts
export { default } from "../../role-persona/src/extensions/pi/adapter.ts";

• Registers tools: memory, knowledge, role_info
• Registers commands: /role, /memories, /memory-log, /memory-fix, /memory-tidy, /memory-tidy-llm, /memory-vector, /memory-export, /memory-conflicts, /memory-distill, /memory-distill-stop, /memory-tags, /kb
• Handles events: session_start, before_agent_start, agent_end, session_before_compact, session_shutdown, turn_end
• Uses Pi SDK directly for LLM calls (no CLI subprocess)
• On-demand memory search on first message (vector + keyword hybrid)
• Vector auto-recall injection into system prompt
• External readonly memory hints (configurable)
• Compaction memory extraction (<memory> block parsing)
• Interactive memory→knowledge distillation mode
• TUI role selector and memory viewer (when available)
• HTTP memory server for browser-based browsing
• Auto-repair and pending expiration on role activation

MCP Server

# Start on default port (3939)
bun src/transport/mcp-server.ts

# Custom port
MCP_PORT=8080 bun src/transport/mcp-server.ts

Streamable HTTP transport at /mcp.

Configuration

Config file: ~/.pi/roles/pi-role-persona.jsonc

{
  "autoMemory": {
    "enabled": true,
    "model": [
      { "provider": "nvidia", "model": "deepseek-ai/deepseek-v4-flash" },
      { "provider": "modelscope", "model": "ZhipuAI/GLM-5" }
    ],
    "batchTurns": 5,
    "intervalMs": 1800000
  },
  "vectorMemory": {
    "enabled": true,
    "provider": "openai",
    "model": "text-embedding-3-small",
    "autoRecall": true,
    "hybridSearch": true
  },
  "knowledge": {
    "enabled": true
  },
  "logging": {
    "enabled": true,
    "level": "info"
  }
}

Config can be edited via:

• Web UI Form mode — schema-driven form with dropdowns for model selection (reads models.json)
• Web UI JSONC mode — raw editor with jsonc-parser (comments preserved on edit)
• CLI — role-persona commands

Environment Variables

Daemon HTTP API

All responses: { "ok": true, "data": ... } or { "ok": false, "error": { "code": "...", "message": "..." } }

Project Structure

role-persona/
├── src/
│   ├── core/                    # Pure logic, zero Pi dependency
│   │   ├── types.ts             # Shared types (ModelRegistry, LlmCaller, ApiKeyResolver...)
│   │   ├── config.ts            # JSONC config loader
│   │   ├── logger.ts            # JSONL structured logging
│   │   ├── role-store.ts        # Role CRUD, CWD mapping
│   │   ├── memory-md.ts         # Markdown memory CRUD
│   │   ├── memory-llm.ts        # LLM auto-extraction
│   │   ├── memory-tags.ts       # LLM tag extraction
│   │   ├── memory-vector.ts     # Vector search orchestration
│   │   ├── embedding.ts         # Embedding providers (OpenAI/Local/MiniLM)
│   │   ├── vector-db.ts         # LanceDB wrapper
│   │   └── knowledge.ts         # Knowledge base CRUD
│   ├── service/                 # Unified facade
│   │   ├── context.ts           # ServiceContext
│   │   ├── index.ts             # RolePersonaService
│   │   └── *-service.ts         # Sub-services
│   ├── extensions/pi/
│   │   └── adapter.ts           # Pi extension (direct service + Pi SDK)
│   ├── transport/
│   │   ├── daemon.ts            # HTTP daemon (ServiceManager + multiplexing + static Web UI)
│   │   ├── mcp-server.ts        # MCP Streamable HTTP
│   │   └── tui-renderers.ts     # Pi TUI renderers
│   └── bin/
│       ├── cli.ts               # CLI entry (--web, --cwd, --direct)
│       └── daemon.ts            # Daemon entry
├── tests/
│   ├── memory-md.test.ts        # Memory CRUD tests (26/26)
│   ├── core.test.ts             # Core types/config tests
│   ├── cli.test.ts              # CLI integration tests
│   └── mcp.test.ts              # MCP protocol tests
├── web/                         # React dashboard
│   ├── src/
│   │   ├── api/client.ts        # Daemon API client (role-based)
│   │   ├── i18n/                # en.json, zh.json
│   │   ├── hooks/
│   │   │   ├── useApi.ts        # API call hook with loading/error state
│   │   │   └── useUrlState.ts   # URL search param persistence
│   │   ├── components/
│   │   │   ├── Layout.tsx       # Sidebar + topbar + role selector
│   │   │   ├── MarkdownViewer.tsx # Markdown + syntax highlight + edit
│   │   │   ├── JsoncForm.tsx    # Schema-driven JSONC form editor
│   │   │   ├── StatCard.tsx     # Dashboard stat card
│   │   │   └── EmptyState.tsx   # Empty state placeholder
│   │   └── pages/
│   │       ├── Dashboard.tsx    # Stats overview with clickable cards
│   │       ├── Memory.tsx       # Explorer tree + table + Markdown
│   │       ├── Knowledge.tsx    # List + search + Markdown view
│   │       ├── Roles.tsx        # Role grid + detail modal
│   │       └── Settings.tsx     # Config editor + daemon status
│   └── package.json
├── templates/
│   └── memory-export.html       # HTML export template
└── README.md

Memory System

Types

Lifecycle

Conversation → Auto-Extract → Pending Layer → Verify → Consolidated
                                    ↓
                              Use-driven value

Vector Memory

• LanceDB vector storage
• Hybrid search: vector + keyword → RRF fusion
• Auto-index on write
• Auto-recall at session start
• Embedding providers: OpenAI, local, MiniLM (ONNX)

Scenario Memory (L2)

Scenario memory records “what to do in a recurring situation”. It sits between pending atomic facts and consolidated persona guidance, and is useful for code review formats, release procedures, debugging SOPs, and preferred report structures.

role-persona memory scenario-write \
  --title "Code review output" \
  --guidance "Give the conclusion first, then group findings by Critical/Medium/Low. Explain impact and fix for each item." \
  --triggers "code review,review feedback" \
  --evidence "User prefers structured severity-based review feedback"

role-persona memory scenario-search "review this PR"

When building a prompt, role-persona performs on-demand recall from the user query. Matching scenarios are injected as Scenario Memory Hints. They are hints, not commands, and never override explicit user instructions.

Fusion design

If you are looking at Tencent's newly discussed auto-memory approach, start here:

• docs/MEMORY-FUSION-DESIGN.md — maps TencentDB Agent Memory's layered long-term memory, symbolic short-term memory, and traceable drill-down model onto role-persona's daily / pending / consolidated / vector system.

Logging

JSONL structured logging: ~/.pi/roles/.log/YYYY-MM-DD.jsonl

{"schema":"2.0.0","timestamp":"2026-05-07T00:16:00Z","level":"info","scope":"auto-extract","message":"start"}

Development

bun run typecheck        # Type check
bun run test             # All tests (57 tests, 55 pass)
bun test tests/memory-md.test.ts  # Memory module tests (26/26)
bun src/bin/cli.ts ...   # CLI dev
bun src/transport/daemon.ts  # Daemon dev
cd web && bun run dev    # Web dev (Vite HMR)

Test Coverage

License

MIT © Dwsy