Telemem

What telemem does

TeleMem provides persistent memory for agents operating across multiple conversations or runs. The MCP interface accepts either a single statement or a sequence of conversation messages, then stores the resulting memory in a selected scope. Scopes can identify a user, agent, or run; when no user is supplied for ordinary operations, the server uses its configured default user.

Semantic retrieval is handled by search_memories. Results are returned as one consolidated passage rather than as unrelated individual matches, because TeleMem fuses related memories. Shared conversation-event memories use the reserved pseudo-user events and are included automatically in searches. The underlying project also targets character-aware memory, multimodal reasoning, and video processing, alongside its text-memory workflow.

How it works

TeleMem's memory pipeline extracts important information from dialogue, groups semantically similar content, removes or merges duplicates, and stores searchable data in a local FAISS index with JSON metadata. Retrieval combines vector search with reranking. A write buffer supports batched persistence, while an LLM-based clustering and fusion stage helps consolidate related memories.

The MCP layer exposes the memory lifecycle directly. Agents can add or search content, enumerate memories and their identifiers, fetch one memory, replace its text, remove one memory, clear a specifically named scope, or inspect change events. History records ADD, UPDATE, and DELETE operations for an individual memory.

Setup and configuration

The package is available through PyPI and can be installed with pip install telemem. The MCP registry also supports starting it without a separate installation by running uvx telemem. TeleMem is designed to run locally by default, using Qwen and FAISS rather than sending memory data to a cloud service.

The project provides a Mem0-compatible API, including matching add() and search() argument and result shapes, so applications using that interface can use TeleMem as a replacement outside the MCP tool layer. The repository also includes configuration options for Ollama, DeepSeek, and Kimi, as well as examples for LangChain and LlamaIndex.

Tools and capabilities

• add_memory stores a fact, preference, or conversation using text or messages.
• search_memories performs semantic search and returns a consolidated text passage.
• get_memories lists memories and their IDs for a user, agent, or run; events lists shared event memories.
• get_memory retrieves one memory by ID.
• update_memory replaces the text associated with an existing ID.
• delete_memory removes one memory.
• delete_all_memories clears a scope only when an explicit user, agent, or run identifier is supplied.
• memory_history reports a memory's add, update, and delete events.

Limitations and notes

Deleting all memories is intentionally stricter than ordinary operations: the tool requires an explicit scope and never assumes the default user. The local deployment model keeps data on the host, but the memory pipeline can use language models for extraction and semantic clustering according to the selected configuration. Video and multimodal processing are project capabilities, while the listed MCP interface centers on memory records and conversation content.