TeleAI-UAGI/telemem

What it solves

TeleMem addresses the need for AI agents to maintain consistent, long-term memory across complex, multi-turn interactions. It specifically solves the problem of "memory confusion" in multi-character environments and provides a way for agents to store and reason over multimodal content like videos, which standard memory systems often struggle to handle.

How it works

TeleMem acts as a memory management layer that sits between your AI agent and its storage. It processes incoming dialogue using a pipeline that includes character-aware summarization, semantic clustering to merge similar memories, and efficient batch writing. It stores data using a dual-system approach: FAISS for fast vector-based retrieval and JSON for human-readable metadata. For video, it extracts frames and generates captions to build a searchable vector database, allowing agents to perform multi-step reasoning over video content.

Who it’s for

It is designed for developers building virtual agents, AI companions, NPCs, or customer service bots that require personalized, long-term memory. It is also ideal for those who need a high-performance, drop-in replacement for existing memory systems or want to run their memory infrastructure entirely locally for privacy.

Highlights

• Drop-in compatibility: Works as a high-performance replacement for Mem0 with a single line of code.
• Character-aware: Automatically builds isolated, independent memory profiles for different characters to prevent data mixing.
• Multimodal support: Includes a full pipeline to process, store, and reason over video content using ReAct-style multi-step QA.
• Local-first: Capable of running entirely on your own hardware (using Qwen and FAISS) with no cloud dependency.
• Performance-optimized: Features batch-flushing and concurrency for 2–3x faster write speeds compared to standard streaming methods.