telemem

TeleMem: Building Long-Term and Multimodal Memory for Agentic AI

TeleMem is an agent memory management layer that can be used as a high-performance drop-in replacement for Mem0 with one line of code (import telemem as mem0), deeply optimized for complex scenarios involving multi-turn dialogues, character modeling, long-term information storage, and semantic retrieval.

Through its unique context-aware enhancement mechanism, TeleMem provides conversational AI with core infrastructure offering higher accuracy, faster performance, and stronger character memory capabilities.

Building upon this foundation, TeleMem implements video understanding, multimodal reasoning, and visual question answering capabilities. Through a complete pipeline of video frame extraction, caption generation, and vector database construction, AI Agents can effortlessly store, retrieve, and reason over video content just like handling text memories.

The ultimate goal of the TeleMem project is to use an agent’s hindsight to improve its foresight.

TeleMem, where memory lives on and intelligence grows strong.

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 + 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.

📢 Latest Updates

• [2026-08-15] 🎉 TeleMem v1.10.0 adds first-class DeepSeek Harness support: an opt-in Cordis/MCP patch starts TeleMem with uvx, registers all 8 memory tools under mcp__telemem__*, and securely forwards provider configuration. See the MCP server docs.
• [2026-08-06] 🎉 TeleMem v1.9.0 is on the latest MCP standard: migrated to the official MCP Python SDK v2 (spec 2026-07-28) — all 8 tools now declare titles, behavior annotations, and structured output, while staying compatible with older MCP clients. See the MCP server docs.
• [2026-07-11] 🎉 TeleMem v1.8.0 — the “claims = contracts” release: character-memory extraction fix, infer=False/prompt/memory_type now fully honored, offline contract test suite, telemetry disabled by default, and a multi-NPC demo!
• [2026-06-12] 🎉 TeleMem v1.7.1 is live on the official MCP registry — run the memory server with zero install: uvx telemem! Also new: evaluation principles and a LongMemEval harness with built-in baselines.
• [2026-06-12] 🎉 TeleMem is now on PyPI: pip install telemem! v1.6.0 adds Ollama/DeepSeek/Kimi configs, LangChain & LlamaIndex examples, and a documentation site.
• [2026-06-12] 🎉 TeleMem v1.5.0 has been released: true mem0 drop-in API, lightweight core install, and CI!
• [2026-06-11] 🎉 TeleMem v1.4.0 has been released with MCP support!
• [2026-01-28] 🎉 TeleMem v1.3.0 has been released!
• [2026-01-22] 🎉 TeleMem Tech Report has been updated to its 4th version!
• [2026-01-13] 🎉 TeleMem Tech Report has been released on arXiv!
• [2026-01-09] 🎉 TeleMem v1.2.0 has been released!
• [2025-12-31] 🎉 TeleMem v1.1.0 has been released!
• [2025-12-05] 🎉 TeleMem v1.0.0 has been released!

🔥 Research Highlights

• Significantly improved memory accuracy: Achieved 86.33% accuracy on the ZH-4O Chinese multi-character long-dialogue benchmark, 19% higher than Mem0.
• Doubled speed performance: Millisecond-level semantic retrieval enabled by efficient buffering and batch writing.
• Greatly reduced token cost: Optimized token usage delivers the same performance with significantly lower LLM overhead.
• Precise character memory preservation: Automatically builds independent memory profiles for each character, eliminating confusion.
• Automated Video Processing Pipeline: From raw video → frame extraction → caption generation → vector database, fully automated
• ReAct-Style Video QA: Multi-step reasoning + tool calling for precise video content understanding

📌 Table of Contents

• Project Introduction
• TeleMem vs Mem0: Core Advantages
• Experimental Results
• Quick Start
• Project Structure
• Core Functions
• Multimodal Extensions
• MCP Server
• Framework Integrations
• Data Storage Explanation
• Telemetry & Privacy
• Development and Contribution
• Acknowledgements
• Citation

Project Introduction

TeleMem enables conversational AI to maintain stable, natural, and continuous worldviews and character settings during long-term interactions through a deeply optimized pipeline of character-aware summarization → semantic clustering deduplication → efficient storage → precise retrieval.

Features

• Automatic memory extraction: Extracts and structures key facts from dialogues.
• Semantic clustering & deduplication: Uses LLMs to semantically merge similar memories, reducing conflicts and improving consistency.
• Character-profiled memory management: Builds independent memory archives for each character in a dialogue, ensuring precise isolation and personalized management.
• Efficient asynchronous writing: Employs a buffer + batch-flush mechanism for high-performance, stable persistence.
• Precise semantic retrieval: Combines FAISS + JSON dual storage for fast recall and human-readable auditability.

Applicable Scenarios

• Multi-character virtual agent systems

• Long-memory AI assistants (e.g., customer service, companionship, creative co-pilots)

• Complex narrative/world-building in virtual environments

• Dialogue scenarios with strong contextual dependencies

• Video content QA and reasoning

• Multimodal agent memory management

• Long video understanding and information retrieval

  

TeleMem vs Mem0: Core Advantages

TeleMem deeply refactors Mem0 to address characterization, long-term memory, and high performance. Key differences:

Experimental Results

Dataset

We evaluate the ZH-4O Chinese long-character dialogue dataset constructed in the paper MOOM: Maintenance, Organization and Optimization of Memory in Ultra-Long Role-Playing Dialogues:

• Average dialogue length: 600 turns per conversation
• Scenarios: daily interactions, plot progression, evolving character relationships

Memory capability was assessed via QA benchmarks, e.g.:

{
"question": "What is Zhao Qi's nickname for Bai Yulan? A Xiaobai B Xiaoyu C Lanlan D Yuyu",
"answer": "A"
},
{
"question": "What is the relationship between Zhao Qi and Bai Yulan? A Classmates B Teacher and student C Enemies D Neighbors",
"answer": "B"
}

Experimental Configuration

• LLM: Qwen3-8B (thinking mode disabled)

• Embedding model: Qwen3-Embedding-8B

• Metric: QA accuracy

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

Quick Start

Installation

pip install telemem            # core (text memory)
pip install "telemem[mcp]"     # + MCP server
pip install "telemem[video]"   # + video/multimodal pipeline
pip install "telemem[all]"     # everything

Development Environment

Using uv (recommended — creates .venv from the committed uv.lock for a reproducible environment):

uv sync --all-extras   # install TeleMem (editable) + all extras, incl. MCP
uv run python examples/quickstart.py

Or with conda + pip:

# Create and activate virtual environment
conda create -n telemem python=3.10
conda activate telemem
# Install from source (editable), with the extras you need
pip install -e ".[all]"

Example

Set your OpenAI API key:

export OPENAI_API_KEY="your-openai-api-key"

# python examples/quickstart.py
import telemem as mem0

memory = mem0.Memory()

messages = [
    {"role": "user", "content": "Jordan, did you take the subway to work again today?"},
    {"role": "assistant", "content": "Yes, James. The subway is much faster than driving. I leave at 7 o'clock and it's just not crowded."},
    {"role": "user", "content": "Jordan, I want to try taking the subway too. Can you tell me which station is closest?"},
    {"role": "assistant", "content": "Of course, James. You take Line 2 to Civic Center Station, exit from Exit A, and walk 5 minutes to the company."}
]

memory.add(messages=messages, user_id="Jordan")
results = memory.search("What transportation did Jordan use to go to work today?", user_id="Jordan")
for hit in results["results"]:   # same result shape as mem0
    print(hit["memory"])

Memory() uses the default provider settings inherited from mem0ai. To use the repository’s local Qwen + FAISS configuration, load config/config.yaml explicitly:

from telemem.utils import load_config
import telemem as mem0

config = load_config("config/config.yaml")
memory = mem0.Memory(config=config)

The runnable examples also honor the same configuration through TELEMEM_CONFIG:

TELEMEM_CONFIG=config/config.yaml python examples/quickstart.py

Using MiniMax as the LLM Provider

TeleMem supports MiniMax as an LLM backend via its OpenAI-compatible API. A ready-to-use example config is provided at config/config.minimax.yaml.

export MINIMAX_API_KEY="your-minimax-api-key"
export OPENAI_API_KEY="your-openai-api-key"  # still needed for embeddings

from telemem.utils import load_config
import telemem as mem0

config = load_config("config/config.minimax.yaml")
memory = mem0.Memory(config=config)

Key points for MiniMax usage:

• LLM: MiniMax-M3.1-Flash (default) or MiniMax-M3, both with a 1,000,000-token context window and text/image/video input. Existing configs keep their explicit model IDs, including legacy and custom models.
• Regional endpoints: use https://api.minimax.io/v1 (global) or https://api.minimax.cn/v1 (China) as openai_base_url; existing custom endpoint values are preserved.
• Thinking: Flash always uses adaptive thinking, with max reasoning effort when omitted. The preset sets is_reasoning_model: true so mem0 forwards an optional reasoning_effort (low, medium, high, xhigh, or max) and omits its small default token cap. Flash rejects disabled thinking, reasoning_effort="none", service_tier, and reasoning_split=false. See the OpenAI-compatible API contract for details.
• Tool history: when calling the LLM adapter directly, use mem0’s response_callback to capture the complete SDK assistant message and append it with tool results on the next call. Preserve both reasoning_content and tool_calls; the adapter’s simplified return value is not a complete assistant message.
• Temperature: ignored by Flash and M3, which use a fixed value of 1.
• Embeddings: MiniMax does not provide a public embedding API; configure a separate embedder (e.g. text-embedding-3-small) in the embedder section

More LLM Providers

TeleMem works with any OpenAI-compatible endpoint. Ready-to-use config examples ship in config/:

TELEMEM_CONFIG=config/config.ollama.yaml python examples/quickstart.py   # 100% local memory

Project Structure

telemem/
├── assets/                 # Documentation assets and figures
├── baselines/              # Baseline implementations for comparative evaluation
│ ├── RAG                   # Retrieval-Augmented Generation baseline
│ ├── MemoBase              # MemoBase memory management system
│ ├── MOOM                  # MOOM dual-branch narrative memory framework
│ ├── A-mem                 # A-mem agent memory baseline
│ └── Mem0                  # Mem0 baseline implementation
├── config/               
│ ├── config.yaml           # TeleMem default configuration
│ └── config.minimax.yaml   # MiniMax provider example configuration
├── data/                   # Small sample datasets for evaluation or demonstration
├── examples/               # Code examples and tutorial demos
│ ├── quickstart.py         # Quick start
│ ├── quickstart_mm.py      # Quick start (multimodal)
│ ├── mcp_client.py         # Quick start over MCP (stdio client)
│ ├── mcp_config.json       # MCP config snippet for Claude Desktop / Cursor
│ └── deepseek-harness.cordis.yml # DeepSeek Harness memory patch
├── docs/
│ ├── MCP.md                # MCP server reference
│ └── TeleMem_Tech_Report.pdf
├── telemem/                # Telemem code
│ └── mcp/                  # Model Context Protocol server
├── tests/                  # Telemem test
├── README.md               # English README
├── README-ZH.md            # Chinese README
└── pyproject.toml          # Python environment

Core Functions

Add Memory (add)

The add() method injects one or more dialogue turns into the memory system.

def add(
 self,
 messages,
 *,
 user_id: Optional[str] = None,
 agent_id: Optional[str] = None,
 run_id: Optional[str] = None,
 metadata: Optional[Dict[str, Any]] = None,
 infer: bool = True,
 memory_type: Optional[str] = None,
 prompt: Optional[str] = None,
 batch: bool = False,
)

🔎 Parameter Description

Returns the mem0-compatible shape: {"results": [{"id": "...", "memory": "...", "event": "ADD"}, ...]}

🔁 Internal Workflow of add()

1. Message preprocessing: Merge consecutive messages from the same speaker; normalize turn structure.
2. Multi-perspective summarization:
   • Global event summary
   • Character 1’s perspective (actions, preferences, relationships)
   • Character 2’s perspective
3. Vectorization & similarity search: Generate embeddings and retrieve existing similar memories.
4. Batch processing: When buffer threshold is reached, invoke LLM to semantically merge similar memories.
5. Persistence: Dual-write to FAISS (for retrieval) and JSON (for metadata).

🎭 Multi-character demo: examples/multi_npc.py runs five tavern NPCs through one scene — a single add_batch(scene, user_id=[...]) call gives each NPC a private memory profile plus a shared "events" world-state, and each NPC then recalls the scene from their own perspective.

Search Memory (search)

Performs semantic vector-based retrieval of relevant memories with context-aware recall.

def search(
 self,
 query: str,
 *,
 user_id: Optional[str] = None,
 agent_id: Optional[str] = None,
 run_id: Optional[str] = None,
 limit: int = 100,
 filters: Optional[Dict[str, Any]] = None,
 threshold: Optional[float] = None,
 rerank: bool = True,
)

🔎 Parameter Description

Returns the mem0-compatible shape: {"results": [{"id": "...", "memory": "...", "score": ..., ...}, ...]}

🔍 Search is based on FAISS vector retrieval, supporting millisecond-level responses.

Multimodal Extensions

Beyond text memory, TeleMem further extends multimodal capabilities. Drawing inspiration from Deep Video Discovery’s Agentic Search and Tool Use approach, we implemented two core methods in the TeleMemory class to support intelligent storage and semantic retrieval of video content.

Add Multimodal Memory (add_mm)

def add_mm(
    self,
    video_path: str,
    output_dir: str,
    clip_secs: int | None = None,
    emb_dim: int | None = None,
    subtitle_path: str | None = None,
)

🔎 Parameter Description

🔁 add_mm() Internal Flow

1. Frame Extraction: decode_video_to_frames - Decodes video to JPEG frames at configured FPS
2. Caption Generation: process_video - Uses VLM (e.g., Qwen3-Omni) to generate detailed descriptions for each clip
3. Vector Database Construction: init_single_video_db - Generates embeddings for semantic retrieval

💡 Smart Caching: If the target file for a stage already exists, that stage is automatically skipped to save computational resources.

Return Value Example

{
    "output_dir": "/abs/path/to/output_dir"
}

Search Multimodal Memory (search_mm)

def search_mm(
    self,
    question: str,
    output_dir: str,
    max_iterations: int = 15,
)

🔎 Parameter Description

🛠️ ReAct-Style Reasoning Tools

search_mm internally uses MMCoreAgent, employing a THINK → ACTION → OBSERVATION loop with three specialized tools:

Multimodal Example

Run the multimodal demo:

python examples/quickstart_mm.py

On the first run, frames, captions and VDB JSON will be generated under the chosen output_dir. The repository ships a small sample video; generating captions and the video database still requires configured VLM and embedding services unless you already have these artifacts locally.

Complete code example:

import telemem as mem0
from pathlib import Path
from telemem.mm_utils.core import extract_choice_from_msg

# Initialize
memory = mem0.Memory()

# Define paths
repo_root = Path(__file__).resolve().parents[1]
video_path = repo_root / "data" / "samples" / "video" / "3EQLFHRHpag.mp4"
video_name = video_path.stem
output_dir = video_path.parent


# Step 1: Add video to memory (auto-processing)
vdb_json_path = output_dir / "vdb" / video_name / f"{video_name}_vdb.json"
if not vdb_json_path.exists():
    result = memory.add_mm(
        video_path=str(video_path),
        output_dir=str(output_dir),
    )
    print(f"Video processing complete: {result}")
else:
    print(f"VDB already exists: {vdb_json_path}")

# Step 2: Query video content
question = """The problems people encounter in the video are caused by what?
(A) Catastrophic weather.
(B) Global warming.
(C) Financial crisis.
(D) Oil crisis.
"""

messages = memory.search_mm(
    question=question,
    output_dir=str(output_dir),
    max_iterations=15,
)

# Extract final answer
answer = extract_choice_from_msg(messages)
print(f"Answer: ({answer})")

MCP Server

TeleMem ships a Model Context Protocol (MCP) server, so any MCP-compatible client — Claude Desktop, Claude Code, Cursor, custom agents — can use TeleMem as its long-term memory.

pip install telemem

telemem-mcp                                      # stdio (default)
telemem-mcp --transport streamable-http          # Streamable HTTP on :8421
TELEMEM_CONFIG=config/config.yaml telemem-mcp    # custom TeleMem config
uvx telemem                                      # zero-install run (stdio)

Built on the official MCP Python SDK v2, the server implements the current MCP specification (2026-07-28) while remaining compatible with older clients; every tool declares titles, behavior annotations (read-only/destructive hints), and structured output.

The server exposes eight tools: add_memory, search_memories, get_memories, get_memory, update_memory, delete_memory, delete_all_memories, and memory_history. Calls without an explicit scope default to TELEMEM_DEFAULT_USER_ID (telemem-mcp); destructive bulk deletion always requires an explicit scope.

Claude Desktop / Cursor configuration (examples/mcp_config.json):

{
  "mcpServers": {
    "telemem": {
      "command": "telemem-mcp",
      "env": {
        "TELEMEM_CONFIG": "/absolute/path/to/config/config.yaml",
        "OPENAI_API_KEY": "sk-..."
      }
    }
  }
}

Or drive it programmatically over stdio — the quickstart flow as MCP tool calls:

python examples/mcp_client.py

See docs/MCP.md for the full tool reference, transports, and client setup.

Framework Integrations

TeleMem drops into any agent framework with the same two calls — search() before answering, add() after each exchange:

Because TeleMem is mem0 API-compatible, any framework adapter written for Mem0’s OSS client also works — point it at telemem.Memory instead.

Data Storage

Text Memory Storage

TeleMem automatically creates a structured storage layout under ./faiss_db/, organized by session and character:

faiss_db/
├── session_001_events.index
├── session_001_events_meta.json 
├── session_001_person_1.index 
├── session_001_person_1_meta.json 
├── session_001_person_2.index 
└── session_001_person_2_meta.json 

📄 Metadata Example (_meta.json)

{
 "summary": "Characters discussed the upcoming action plan.",
 "sample_id": "session_001",
 "round_index": 3,
 "timestamp": "2024-01-01T00:00:00Z",
 "user": "Jordan" // Only present in person_*.json
}

All memories include summary, round number, timestamp, and character, facilitating auditing and debugging.

Multimodal Memory Storage

TeleMem generates video-related storage files in the .data/samples/video/ directory:

video/
├── frames/
│   └── <video_name>/
│       └── frames/
│           ├── frame_000001_n0.00.jpg
│           ├── frame_000002_n0.50.jpg
│           └── ...
├── captions/
│   └── <video_name>/
│       ├── captions.json          # Clip descriptions + subject registry
│       └── ckpt/                  # Checkpoint for resume
│           ├── 0_10.json
│           └── 10_20.json
└── vdb/
    └── <video_name>/
        └── <video_name>_vdb.json  # Semantic retrieval vector database

📄 captions.json Structure

{
    "0_10": {
        "caption": "The narrator discusses climate data, showing melting glaciers..."
    },
    "10_20": {
        "caption": "Scene shifts to coastal communities affected by rising sea levels..."
    },
    "subject_registry": {
        "narrator": {
            "name": "narrator",
            "appearance": ["professional attire"],
            "identity": ["climate scientist"],
            "first_seen": "00:00:00"
        }
    }
}

Telemetry & Privacy

TeleMem itself collects no telemetry. The underlying mem0ai library ships anonymized PostHog usage telemetry, which TeleMem disables by default (import telemem sets MEM0_TELEMETRY=False unless you have already set it). To opt back in:

export MEM0_TELEMETRY=true

Development and Contribution

• Issues and pull requests are welcome — see the Contributing Guide to get started.
• Changes between releases are tracked in the Changelog.
• CI runs the offline test suite (uv run pytest tests/ -q) on Python 3.10–3.12 for every PR.
• Chinese documentation: README-ZH.md

License

Apache 2.0 License

Acknowledgements

TeleMem’s development has been deeply inspired by open-source communities and cutting-edge research. We extend our sincere gratitude to the following projects and teams:

• Mem0
• Memobase
• MOOM
• DVD
• Memento
• Momento-Skills

📑 Citation

If you find TeleMem useful to your research or development, please cite our arXiv paper:

Chunliang Chen, Ming Guan, Xiao Lin, Jiaxu Li, Luxi Lin, Qiyi Wang, Xiangyu Chen, Jixiang Luo, Changzhi Sun, Dell Zhang, Xuelong Li. TeleMem: Building Long-Term and Multimodal Memory for Agentic AI. arXiv:2601.06037, 2026. https://arxiv.org/abs/2601.06037

@article{chen2026telemem,
  title   = {TeleMem: Building Long-Term and Multimodal Memory for Agentic AI},
  author  = {Chen, Chunliang and Guan, Ming and Lin, Xiao and Li, Jiaxu and Lin, Luxi and Wang, Qiyi and Chen, Xiangyu and Luo, Jixiang and Sun, Changzhi and Zhang, Dell and Li, Xuelong},
  journal = {arXiv preprint arXiv:2601.06037},
  year    = {2026},
  url     = {https://arxiv.org/abs/2601.06037}
}

Citation metadata is also available in CITATION.cff (GitHub’s “Cite this repository” button).

   

_{mcp-name: io.github.TeleAI-UAGI/telemem}