TeleAI-UAGI/telemem

What it solves

TeleMem is an agent memory management layer designed to solve the problem of maintaining stable, long-term, and character-consistent memories for AI agents. It addresses common issues in conversational AI such as memory confusion between different characters, slow write performance, and the inability to reason over multimodal (specifically video) content.

How it works

TeleMem uses a pipeline of character-aware summarization, semantic clustering, and efficient storage. When dialogue messages are added, the system generates summaries from multiple perspectives (global and per-character). It then uses a buffer and batch-flush mechanism to semantically merge similar memories using an LLM, reducing redundancy. For storage, it employs a dual-write system using FAISS for fast vector retrieval and JSON for human-readable metadata. For video, it implements a pipeline that extracts frames, generates captions, and stores them in a vector database to enable ReAct-style multi-step reasoning over video content.

Who it’s for

This project is intended for developers building:

• Multi-character virtual agent systems or NPCs.
• Long-memory AI assistants for companionship or customer service.
• Complex narrative or world-building environments.
• Agents that require the ability to store and reason over video content.

Highlights

• Mem0 Compatibility: Acts as a high-performance drop-in replacement for Mem0 with a compatible API.
• Character-Isolated Memory: Automatically builds independent memory profiles for each character to prevent persona confusion.
• Multimodal Video Memory: Full pipeline for video-to-text memory conversion and multi-step video QA.
• Local-First: Supports fully local execution using Qwen and FAISS, with integrations for Ollama, DeepSeek, and MiniMax.
• High Performance: Features batch-flush concurrency for faster writes and improved accuracy on long-dialogue benchmarks.