Five-minute quickstart

1. Create an account

MemHQ is free to start — no credit card. Head to sign-up and create an account. We'll auto-provision a personal organization the first time you land in the dashboard.

2. Create a project & grab a key

The first time you open the dashboard, the onboarding flow pops a "create project" modal. Name it whatever you like — my-agent, billing-bot, anything. When the project is created the API key appears once in the modal. Copy it now — it isn't stored plaintext after that.

Stash the key in your environment:

export MEMHQ_API_KEY=mem_xxxxxxxxxxxxxxxxxxxx
# The SDK defaults to https://api.memhq.ai — no other config required.

3. Install the SDK

pip install memhq

4. Add your first memory

One call. Auto-creates the user, auto-creates a default thread, queues every user-role message for extraction. Returns immediately — extraction runs in the background.

import os
from memhq import MemoryClient

client = MemoryClient(api_key=os.environ["MEMHQ_API_KEY"])

client.add(
    messages=[{"role": "user", "content": "I'm a vegetarian, allergic to nuts."}],
    user_id="user_123",
)

5. Search

Hybrid retrieval (BM25 + pgvector + light graph traversal) in a single call. Scope to a user, a set of group graphs, or both.

results = client.search("dietary restrictions", user_id="user_123")
for memory in results:
    print(memory.score, memory.content)

6. Ask a question

This is the MemHQ wedge: a single call that retrieves the relevant memories, reranks them, and synthesizes a cited answer. Drop it straight into your agent's response.

answer = client.ask("What should I avoid eating?", user_id="user_123")
print(answer.text)
for cit in answer.citations:
    print(" •", cit.content)

What just happened

That add-and-ask cycle did a lot of work behind the scenes:

• Auto-provisioning: the SDK's first add() created a User (keyed by your user_id), spun up their personal memory graph, and opened a default thread — all without an explicit setup step.
• Extraction: the worker called the configured LLM (Gemini 2.0 Flash by default — about $0.10 per 1k memories) to pull typed facts out of the message: FACT, PREFERENCE, RELATIONSHIP, plus seven others.
• Hybrid retrieval: search runs BM25 + pgvector + graph traversal in a single SQL query, then reranks. No magic — just well-tuned Postgres.
• Cited synthesis: ask adds an LLM synthesis pass on top of retrieval, so you get an answer plus the specific memories that backed it — no chunk-stuffing required.