mem0

Facts

Openness

Public code, licence not stated

Licence

UNKNOWN

Language

Python

Pinned version

94c3fe9 (2026-09-25)

Topics checked

28 of 33

Kind of agent

General framework

Source repository (opens another site)

What it does notably well

• Concurrency

  mem0 batches embedding (OpenAI embed_batch sends chunks of 100 texts) and insertion (one vector_store.insert for all extracted memories) and bounds parallelism at 4 workers for…

• Eval harness

  mem0 has a CI-gated mocked pytest suite: 101 test_*.py files under tests/ (86 of the 104 .py files use Mock/mock/patch), run by pytest tests/ in ci.yml.

• External memory

  In mem0's default add path (infer=True) facts are extracted in one pass by a single JSON-mode LLM call with no tools, and the pipeline only creates ADD events; it has no UPDATE…

  mem0 recall combines semantic, BM25 and entity-boost scores additively as min((semantic + bm25 + entity_boost) / max_possible, 1.0), where max_possible is 1.0 plus 1.0 if BM25…

  mem0's BM25 and entity signals only re-score candidates that the semantic search returned: the candidate list is built from semantic_results alone and score_and_rank iterates…

• MCP and auth

  mem0 serves MCP from a hand-written stdio JSON-RPC loop in the agent plugins (initialize, ping, tools/list, tools/call) that exposes one read-only tool, search_memories, with a…