Memoryinfrastructure
forAIagents.
Three endpoints turn raw conversation into a queryable, per-user knowledge graph — and stream back answers, not just memories.
- LoCoMo
- 61.0%
- Uptime
- 99.9%
- Endpoints
- 3
Reproducible
Business SLA
add · search · ask
pip install memhq·npm install @memhq/sdk
How it works
Three calls. Add, search, ask.
Most memory libraries hand you a vector store and call it a day. MemHQ gives you the full pipeline — extraction, retrieval, and grounded synthesis — behind a typed API.
Ingest a fact, conversation, or document. An async pipeline extracts memories, embeds them, and stores them in a per-user graph.
mem.add({
userId: "u_42",
content: "Prefers concise replies"
})Hybrid BM25 + vector retrieval with Reciprocal Rank Fusion. Returns ranked memories with scores and provenance.
mem.search({
userId: "u_42",
q: "tone preferences?"
})Server-side: retrieve, ground, and synthesize a cited answer in one round trip. No retrieve-stuff-prompt-LLM dance.
mem.ask({
userId: "u_42",
q: "tone preferences?"
})Use cases
What you build on top of memory.
Chatbots that remember you across sessions
Conversation history compresses into durable memories. Day-30 the bot still knows what day-1 told it.
Agents with long-term memory across runs
State that persists between agent invocations — no scratchpad gymnastics, no re-priming on every loop.
Shared org knowledge, per-user privacy
Group graphs for what the team knows. User graphs for what only they know. Both queried in one call.
RAG-quality answers without building a pipeline
Chunker, vector store, reranker, synthesizer — collapsed into one managed endpoint with citations.
Integrations
Drop into your stack.
Wire MemHQ into the framework you already use. SDKs in Python and TypeScript, an MCP server for desktop and IDE assistants, and a REST API for everything else.
from memhq import MemHQ mem = MemHQ(api_key="sk_live_…") mem.add(user_id="u_42", content="I drive a Rivian R1S and Tesla Model Y") answer = mem.ask(user_id="u_42", q="What cars do I own?") # → "You own a Rivian R1S and a Tesla Model Y." [m_72ad, m_72ae]
Benchmark
Built on a reproducible benchmark, not marketing copy.
MemHQ scores 61.0% on the LoCoMo memory benchmark using standard published methodology — full reproducibility in our docs. Per-category breakdown to the right; the harness, prompts, and snapshots are linked so you can re-run it against your workload.
Start building memory into your agent in 60 seconds.
Free Hobby tier. No card required. Three endpoints between you and an agent that remembers.