TeleMem is an agent memory management layer designed as a drop-in replacement for Mem0, offering enhanced capabilities for long-term dialogue memory, semantic deduplication, and multimodal video reasoning. It targets developers building complex conversational AI, AI assistants, and agents that require robust memory handling, aiming to provide higher accuracy, faster performance, and stronger character memory preservation, while also reducing LLM token costs.

How It Works

TeleMem deeply refactors Mem0, introducing a context-aware enhancement mechanism. Its core approach involves character-aware summarization, LLM-based semantic clustering for deduplication (improving upon Mem0's vector similarity filtering), and efficient asynchronous writing via a buffer-batch-flush system. Data is stored using a dual-write strategy combining FAISS for fast retrieval and JSON for human-readable auditability. For multimodal capabilities, TeleMem implements a full pipeline for video processing, including frame extraction, caption generation, and vector database construction, enabling ReAct-style video question answering and reasoning.

Quick Start & Requirements

• Installation: Create and activate a Conda environment (conda create -n telemem python=3.10, conda activate telemem), then install dependencies (pip install -r requirements.txt). Apply custom patches using bash scripts/apply_patches.sh.
• Prerequisites: An OpenAI API key is required (export OPENAI_API_KEY="your-openai-api-key").
• Documentation: Links to overlay development guides (TeleMem-Overlay.md) and examples (examples/quickstart.py, examples/quickstart_mm.py) are provided.

Highlighted Details

• Achieved 86.33% accuracy on the ZH-4O Chinese multi-character long-dialogue benchmark, a 19% improvement over Mem0.
• Enables millisecond-level semantic retrieval through efficient buffering and batch writing.
• Significantly reduces token costs by optimizing LLM usage.
• Features automated video processing from raw video to a vector database.
• Supports ReAct-style multi-step reasoning for video question answering.
• Automatically builds independent memory profiles for each character in a dialogue.

Maintenance & Community

The project is developed by the Ubiquitous AGI team at TeleAI. It acknowledges inspiration from several open-source projects including Mem0, Memobase, MOOM, and DVD. Specific community links (e.g., Discord, Slack) or a public roadmap are not detailed in the provided README.

Licensing & Compatibility

TeleMem is released under the Apache 2.0 License, which is permissive and generally suitable for commercial use and integration into closed-source projects.

Limitations & Caveats

The project requires an OpenAI API key for its core functionalities. The "overlay" development process, involving applying patches to an upstream repository (Mem0), may introduce complexity in managing updates and future compatibility. The README includes release dates in late 2025, suggesting it may be a relatively new or forward-looking project.