44c332bfe1
新增文件: - README_EN.md (English README) - docs/记忆机制.md + _EN.md (记忆工作机制) - docs/架构_EN.md - docs/一键启动指南_EN.md - docs/工作记忆链_EN.md - docs/api_EN.md - docs/prompts.md + _EN.md docs/README.md 更新为双语文档索引
2.5 KiB
TrulyMEM Documentation
Welcome to the TrulyMEM project documentation.
Documentation Index
- Architecture - System architecture and technical design
- Quick Start - Quick start guide
- Memory Mechanism - Internal memory working mechanism
- Working Memory Chain - Continuous task handling
- BackendServer API - Backend API reference (for extension development)
- Prompt Manager - Prompt management module
Project Introduction
TrulyMEM (TrueHumanMEM) is a graph-based memory system that gives AI long-term memory capabilities, allowing AI to remember, recall, and manage information like humans.
Core Features
- Long-term Memory: SQLite embedded graph database, out-of-the-box
- Persona Graph: Role-playing and character settings support
- Working Memory Chain: Task tracking for conversation continuity
- TUI & Backend Separation: Multi-threaded Queue communication
- Keyboard-driven TUI: Full keyboard operation, no mouse required
- Cross-platform: Windows / Linux / macOS
- Standalone Deployment: Packaged as executable
Story
Industry believes that LLMs' massive parameters give them emergent intelligence. But this intelligence is "dead" — it cannot truly remember, nor understand the concept of "remembering". Everything it outputs is the probabilistic optimal solution calculated through countless forward passes on the current input text. The LLM cannot correct its weights based on errors in a conversation, nor perform backward passes. Its consciousness is frozen — what appears as intelligence is merely the echo of this frozen consciousness.
Current "memory systems" merely externalize memory, letting the "system" remember for the LLM. Or they dump all context text to the LLM. This is a waste of the model's limited input context.
TrulyMEM asks: since the LLM cannot correct model weights in real-time, why not give the memory authority back to the LLM?
We provide a series of mechanisms for the LLM to decide what to remember, what to forget, what's important, what's trivial. The LLM's reasoning process is also its thinking and recalling process. Abandoning the traditional messages array context, all memories are stored as triplets (graph) in the graph database. When the LLM thinks, it can autonomously jump through graph links to associate related relationships, enabling natural association and recall.
Give the LLM true memory.