Mind Memory Lab: Memory-Oriented Agent Framework

Project Overview

Mind Memory Lab is a memory-oriented agent framework laboratory prototype, aiming to validate the feasibility of long-term memory and conversation continuity in LLM applications. Core focus includes: controlled information retrieval, workspace isolation, and turn-level orchestration.

Technical Architecture

Project uses pnpm monorepo organization (9 packages + 1 app):

Package	Responsibility	Tests
`@mind/kernel`	Turn pipeline, domain types, LLM ports	5 passed
`@mind/memory`	Memory persistence, chunking, SQLite FTS indexing	8 passed
`@mind/workspace`	Workspace sandbox and ignore rules	3 passed
`@mind/adapters-llm`	Anthropic / OpenAI / Ollama / Alibaba Cloud adapters	3 passed
`@mind/agent-host`	Multi-turn tool orchestration AgentToolLoop	-
`@mind/tools-core`	Read-only workspace tools and ToolPolicy	-
`@mind/agent-contract`	Agent tools and MCP/Hooks/Skills port types	No runtime deps
`@mind/native-mindmem`	N-API bindings (optional)	Fallback verified
`c/`	C99 primitives (chunking, fingerprinting, counting)	-
`@mind/web`	Web UI + local API	-

Web UI Features

Chat: AI Agent chat interface (session management, streaming response, thinking process display)
Memory Library: Memory item management (list browsing, filtering, detail viewing)
Search: Full-text search functionality

Core Design

Workspace Sandbox

Read-only access control based on workspace sandbox:

interface WorkspaceSandbox {
  rootPath: string
  ignorePatterns: string[]  // gitignore format
  allowPatterns: string[]   // whitelist
  canAccess(path: string): boolean
  readFile(path: string): Promise<string>
}

Memory System

SQLite FTS5 full-text indexing:

interface MemoryStore {
  store(entry: MemoryEntry): Promise<void>
  search(query: string, limit: number): Promise<MemoryEntry[]>
  chunkAndIndex(content: string): Promise<Chunk[]>
}

C99 Primitives

size_t mind_chunk_end(const char* utf8, size_t max_bytes);
uint64_t mind_fnv1a64(const char* data, size_t len);
size_t mind_whitespace_token_estimate(const char* text);

Optional N-API binding to Node.js. Automatic fallback to TypeScript implementation when native module unavailable.

Current Status

Turn pipeline, memory persistence and FTS retrieval, multi-LLM backend adapters, and Web UI implemented. Vector search, message bus, and multi-ThinkingUnit parallelism under development.