Skip to content

TeleMem

Long-term and multimodal memory for agentic AI — a high-performance, mem0-compatible memory layer built by the Ubiquitous AGI team at TeleAI and Bloo-Mind AI.

Tech Report (arXiv) · GitHub · 中文文档

Why TeleMem?

  • 🎭 Character memory done right — the only open-source memory layer that automatically builds isolated, per-character memory profiles, built for role-play, companion AI, NPCs, and multi-persona assistants.
  • 🎬 Memory for video, not just text — a full video → frames → captions → vector DB pipeline with ReAct-style multi-step video QA.
  • 🏠 Fully local by default — runs end-to-end on your hardware (Qwen or Ollama + FAISS); no cloud service, no paid tier, no data leaving your machine.
  • 🔌 mem0-compatible API — add() / search() accept the same arguments and return the same {"results": [...]} shapes, so existing Mem0 code keeps working with import telemem as mem0.

Results

On the ZH-4O Chinese multi-character long-dialogue benchmark (600-turn conversations, QA accuracy):

Method Overall (%)
RAG 62.45
Mem0 70.20
MOOM 72.60
A-mem 73.78
Memobase 76.78
TeleMem 86.33

LLM: Qwen3-8B, embeddings: Qwen3-Embedding-8B. Reproduction harness in baselines/.

How it works

flowchart LR
    A["Dialogue<br/>messages"] --> B["Character-aware<br/>summarization<br/>(global + per-character)"]
    B --> C["Embedding +<br/>similar-memory<br/>retrieval"]
    C --> D["Write buffer<br/>(batch flush)"]
    D --> E["LLM semantic<br/>clustering & fusion"]
    E --> F[("FAISS index +<br/>JSON metadata")]
    Q["Query"] --> S["Vector search<br/>+ rerank"]
    F --> S
    S --> R["results"]

Install

pip install "telemem @ git+https://github.com/TeleAI-UAGI/telemem.git"

# extras
pip install "telemem[mcp] @ git+https://github.com/TeleAI-UAGI/telemem.git"    # MCP server
pip install "telemem[video] @ git+https://github.com/TeleAI-UAGI/telemem.git"  # video pipeline

Continue with the Quickstart.