One API key, isolated memory per user_id. Your agents get semantic facts, episodic events, and procedural workflows that evolve when they fail — extraction, dedup, and contradiction-handling run server-side.
pip install mengram-ai
from mengram import Mengram
m = Mengram(api_key="om-...")
# each of YOUR users gets an isolated memory store
m.add([{"role": "user", "content": "I prefer email, and my last order arrived damaged"}],
user_id="customer-4812")
# later — any session, any of your agents
m.search("how should I contact this customer?", user_id="customer-4812")
# → prefers email; recent damaged-order incidentREST, Python, and JavaScript SDKs. Same isolation over MCP for tool-native agents.
Pass user_id — every end-user gets isolated facts, events, workflows, and a cognitive profile. One key, thousands of users.
Workflows with version history that evolve from failures: revisions record the violated assumption and derive preconditions your agent checks before trusting a workflow.
Send raw conversation turns — facts, events, and workflows are extracted, deduplicated, and contradiction-checked for you. No prompt engineering.
Subscribe to memory events (updates, contradictions) and drive your own flows when the memory changes.
Multilingual embeddings + reranking — your users can talk to your agent in Russian, Chinese, Spanish; recall works.
Apache 2.0 core. Run it on your infra when compliance demands it; same API surface.
MengramRetriever + chat message history — pip install langchain-mengram
memory=MengramMemory(api_key=...) on any Crew — agents get recall and remember tools.
30 memory tools over streamable HTTP at mengram.io/mcp — Claude, Cursor, or your own MCP client. Listed in the official MCP registry.
Plain HTTP for everything else — full reference.
Free: 40 adds + 200 searches/month — enough to wire up and demo. Production tiers from $19 (10k searches) to $99 (30k searches, unlimited sub-users). Full pricing →