TeleAI-UAGI/telemem

TeleMem is a high-performance drop-in replacement for Mem0, featuring semantic deduplication, long-term dialogue memory, and multimodal video reasoning.

478 ★35 forksPythonUpdated 5d ago

What you need to know

TeleMem — an agent memory management layer and high-performance drop-in replacement for Mem0 (one line: import telemem as mem0), with per-character memory profiles, long-term semantic retrieval, and a video/multimodal memory pipeline.

Install

pip install telemem (core text memory)
pip install "telemem[mcp]" (adds MCP server)
pip install "telemem[video]" (adds video/multimodal pipeline)
pip install "telemem[all]" (everything)
uvx telemem (zero-install MCP server run over stdio)

Usage

  • memory = mem0.Memory(); memory.add(messages=messages, user_id="Jordan"); memory.search("...", user_id="Jordan") — same result shape as mem0
  • MCP server: telemem-mcp (stdio default) or telemem-mcp --transport streamable-http (port 8421); configure via TELEMEM_CONFIG and OPENAI_API_KEY
  • Multimodal: add_mm(video_path, output_dir) then search_mm(question, output_dir) with ReAct-style video QA
  • Works with any OpenAI-compatible endpoint; ready configs for Ollama (fully local), DeepSeek, Kimi (Moonshot), and MiniMax

Key features

  • mem0-compatible API — add()/search() accept the same arguments and return {"results": [...]} shapes
  • Automatically builds isolated per-character memory profiles (role-play, companion AI, NPCs, multi-persona assistants)
  • LLM-based semantic clustering and deduplication merging similar memories; context-aware, character-focused summarization
  • FAISS + JSON dual storage for fast retrieval and human-readable auditability; buffer + batch-flush async writes (2–3x faster than streaming writes)
  • Video memory pipeline — frame extraction → caption generation → vector DB, with ReAct-style multi-step video QA (global_browse_tool / clip_search_tool / frame_inspect_tool)
  • MCP server (official SDK v2, spec 2026-07-28) exposing 8 tools: add_memory, search_memories, get_memories, get_memory, update_memory, delete_memory, delete_all_memories, memory_history

Best for

Building long-term, per-character memory for role-play/companion AI, NPCs, and multimodal agents that need video understanding — fully local if desired.

Caveats

  • Requires an LLM and embedding provider (OpenAI API key, or local Qwen + FAISS via Ollama config); video pipeline needs a configured VLM (e.g. Qwen3-Omni) and embedding service
  • Underlying mem0ai library ships PostHog telemetry, which TeleMem disables by default (MEM0_TELEMETRY=False)
  • Destructive bulk deletion (delete_all_memories) always requires an explicit scope
  • Backed by an arXiv tech report; research-oriented project
Clients: Claude Desktop · Claude Code · Cursor
Documentation ↗

Reviewed 2026-08-07

Topics

agentagent-memoryai-agentscharacter-memoryconversational-aifaissllmlong-term-memorymcpmem0memorymodel-context-protocolmultimodalragsemantic-searchvideo-understanding
View on GitHub ↗
Stars
478★
Forks
35
Language
Python
License
Apache-2.0
Created
2025-12-05
Last push
2026-08-01