TeleMem MCP
Long-term agent memory with per-character profiles and video understanding: an MCP server and a drop-in replacement for Mem0
- 492 ★ on the repository
- Last commit
- Apache-2.0
Extended access
Why this level
- Stores dialogue and facts about people on disk and sends turns to the chosen language model
- The delete_all_memories tool wipes the entire memory of a chosen scope
- With the default settings it sends data to the OpenAI API
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add mcp/telememDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- Stores dialogue and facts about people on disk and sends turns to the chosen language model
- The delete_all_memories tool wipes the entire memory of a chosen scope
- With the default settings it sends data to the OpenAI API
This is third-party code. Review the repository files before installing.
What it does
TeleMem is a memory layer for conversational agents. It takes dialogue turns, extracts facts and stores them so the agent finds the right ones many turns later. Each character in a dialogue gets its own memory profile, so in role-play, multi-hero scenes and companion assistants the facts of different characters do not mix. The server merges similar memories through a language model that clusters them and removes duplicates. Writes go in batches through a buffer, and storage is a FAISS index plus JSON metadata that is easy to read by eye. The interface is compatible with Mem0: after import telemem as mem0 the add() and search() methods take the same arguments and return the same structures. The video module splits a clip into frames, writes captions for segments and builds a vector database. The agent asks questions about the video content, and the answer is built through step-by-step ReAct-style reasoning. The MCP server provides eight tools: add_memory, search_memories, get_memories, get_memory, update_memory, delete_memory, delete_all_memories and memory_history. It runs over stdio or Streamable HTTP, connects to Claude Desktop, Claude Code and Cursor, and starts with no install through uvx telemem.
Who it is for. For developers of agents with long dialogues, role-play characters and video who want Mem0-compatible memory and a local setup.
Good fit when
- A dialogue has several characters and each one's facts must be stored separately
- Your code already uses Mem0 and you want to swap it in without rewriting add() and search() calls
- You want memory on your own machine: the Qwen with FAISS and the Ollama configurations run without a cloud
- The agent must answer questions about video content as well as text
- Claude Desktop or Cursor needs long-term memory that connects with one command through uvx
Not a fit when
- You need memory for code and a repository: TeleMem remembers dialogue and builds no call graph
- You need a shared audit log and permission separation between teams: the project has no such tools
- No local model can run and the OpenAI API is unavailable: the default setup needs an OpenAI key
- You need storage on an external database: data lives in FAISS and JSON files next to the project
Example request
Remember that Mark calls Lena "captain", then remind me how Mark addresses Lena and what he said about the tripMCP
- Transport
- stdio, http, sse
- Authentication
- not required
Environment variables
TELEMEM_CONFIG- Path to a YAML or JSON config file. Without it OpenAI models and a local store are used.
OPENAI_API_KEYsecret- Needed in the default configuration and for embeddings in the MiniMax and DeepSeek variants.
TELEMEM_DEFAULT_USER_ID- Default memory scope when a tool call names no user, agent or run.
Limitations
The default setup needs OPENAI_API_KEY. The OpenAI API is unavailable from Russia, so here take the ready configurations for a local stack: config/config.yaml with Qwen and FAISS, or config/config.ollama.yaml for Ollama. The MiniMax and DeepSeek configurations take embeddings from OpenAI and also need its key: DeepSeek has no embedding API.
Search in the MCP server returns one fused text passage, which is how merging of close memories works; there is no ranked list of rows. Conversation events sit under the pseudo-user events. Bulk deletion requires an explicit user, agent or run. When no scope is given, reads and writes go under TELEMEM_DEFAULT_USER_ID, telemem-mcp by default.
The video module is installed separately: pip install "telemem[video]".
How to disable. Remove the server from the client configuration and uninstall the package with `pip uninstall telemem`. Data lives in the `faiss_db/` directory next to the project and in `.data/samples/video/` for video, and you delete them by hand.
FAQ
How do I connect TeleMem to Claude Desktop or Cursor?
Install `pip install "telemem[mcp]"` and add a telemem server with the command `telemem-mcp` to `claude_desktop_config.json` or `mcp.json`. In env set `TELEMEM_CONFIG` to a config path and `OPENAI_API_KEY` if you do not use a local model. Without installing, the server starts with `uvx telemem`.
How is TeleMem different from Mem0?
The interface is compatible, but TeleMem creates a separate profile for each dialogue character and merges similar memories through a language model. Writes go in batches and data lives in FAISS and JSON. It also handles video. On the Chinese ZH-4O set the authors measured 86.33% accuracy, 19% above Mem0.
Can TeleMem run without a cloud?
Yes. `config/config.yaml` uses Qwen and FAISS on your own hardware, and `config/config.ollama.yaml` works through Ollama and needs no OpenAI key.
Does TeleMem collect telemetry?
The project itself collects none. The mem0ai library it builds on sends anonymous PostHog statistics, and TeleMem turns them off on import. To bring them back, set `MEM0_TELEMETRY=true`.
Which tools does the TeleMem MCP server provide?
Eight: `add_memory`, `search_memories`, `get_memories`, `get_memory`, `update_memory`, `delete_memory`, `delete_all_memories` and `memory_history`. The dangerous `delete_all_memories` requires an explicit user, agent or run and wipes nothing by default.
Discussion
No questions or reviews yet.
Related
MCP serversAn MCP server for long-term agent memory: compresses dialogues into atomic facts, resolves pronouns and anchors relative time to absolute dates
Shared local memory for AI agents: Claude Code, Codex, Cursor and others find the agreements of past sessions in it and record new ones
Membase keeps one shared memory for AI agents that Claude Code, Cursor and Codex read through MCP, an SDK or the membase command
A long-term memory server for agents with a Mem0-compatible API, hybrid search, its own MCP server and Docker deployment.
Local memory for an agent: hooks record facts from conversations and the MCP server brings them back in later sessions from one SQLite file
Long-term memory for Claude Code, Cursor and Codex: the agent saves decisions and finds them again in later sessions
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| TeleMem (this card) | yes | not required | 492 | |
| Mem0 | no | 68,887 | the original memory layer TeleMem builds on: without character profiles or video memory | |
| Graphiti MCP | no | API key | 32,776 | a time-aware knowledge graph in a graph database, while TeleMem keeps vectors in FAISS and JSON |
| MCP Memory Service | no | not required | 2,117 | memory with a knowledge graph and a REST API on your own server, without video processing |
README in short
The documentation describes the pipeline from character-aware dialogue summarization to reranked search, compares the project with Mem0 and reports results on the ZH-4O set. Next come a quick start, LangChain and LlamaIndex examples and the storage file layout. A separate MCP.md lists the tools, transports and client setup.