kldhsh123/Afterglow

基于 RAG（免微调）从聊天记录唤起语气与回忆：让曾经对话的温度，延续为往后的陪伴。 | Zero-finetuning RAG to revive tone and memories from chat logs: Let the warmth of past conversations continue into companionship.

What it solves

Afterglow is designed to recreate the feeling of chatting with a specific person by using their real historical chat logs. It allows users to build a digital companion that mimics the personality, expression style, and shared memories of a real individual without needing to fine-tune the underlying LLM.

How it works

The system uses a Retrieval-Augmented Generation (RAG) pipeline combined with persona management. It processes imported chat logs (from platforms like QQ, WeChat, and Douyin) through a pipeline that cleans, splits, and vectorizes the text.

Key components include:

  • Hybrid Retrieval: A five-way vector recall system that retrieves relevant historical texts, image summaries, and live memories using RRF (Reciprocal Rank Fusion).
  • Persona & State Management: It maintains a persona card, circadian (sleep/wake) profiles, and life states to ensure the AI's responses are contextually appropriate to the time of day and the person's habits.
  • Offline Analysis: Tools like xuwen analyze generate timelines and personality reports from the chat history to further refine the AI's behavior.
  • OpenAI Compatible API: The backend provides an API that allows the system to be connected to various third-party clients or bots (e.g., QQBot, WeiXinBot).

Who it’s for

Individuals who want to preserve the linguistic style and memories of a loved one or a friend, or those looking to create highly personalized AI companions based on real-world interaction data.

Highlights

  • Multi-source Import: Supports various chat export formats including JSON and JSONL from multiple social platforms.
  • Local-First Privacy: Uses LanceDB for local storage of vectors and personas, supporting PII redaction and offline model services.
  • Proactive Messaging: Includes an engine that can initiate conversations based on analyzed personality traits and schedules.
  • Memory Layering: Distinguishes between real human history, new user messages, and AI-generated responses to prevent AI content from polluting the long-term persona.

Related

  • Project
    kamjin3086/chatless

    A local-first desktop AI chat client that supports cloud and local models, featuring local RAG for document-based knowledge retrieval and MCP protocol integration.

  • Project
    Capsize-Games/airunner

    AI Runner is a self‑hosted desktop app that offers a voice‑enabled AI companion with long‑term memory and a layered canvas for local AI‑generated art. All inference (LLM, diffusion, TTS/STT) runs on the user’s machine, with optional Docker or headless API deployments.

  • Project
    Johnson-Durui/Companion-Space

    A local-first companion learning app that combines RAG-based document Q&A with 3D virtual characters for interactive study and review.

  • Project
    pmbstyle/Alice

    A desktop AI companion with voice interaction, system-level tool access, and long-term memory that can be run locally or via cloud APIs.

  • Project
    rag-web-ui/rag-web-ui

    An intelligent dialogue system that allows users to build custom Q&A services by combining their own document knowledge bases with LLMs using Retrieval-Augmented Generation.