Portable memory for AI agents

Portable memory with
a knowledge graph you control.

You set the relations. No LLM cost. No hallucinated links. Agents signal each other in real time. Every change traceable by agent. 10 cards free. Graph from $29/mo.

No credit card. Takes 30 seconds. Cancel anytime.

▶ Watch the demo

Works with

🔗 LangGraph ⚡ Claude Code 🔗 Cursor 🤖 LangChain 👥 CrewAI ⚙️ AutoGPT 🐍 Python 💚 Node.js ⚙️ VS Code 🌎 REST API 🔗 LangGraph ⚡ Claude Code 🔗 Cursor 🤖 LangChain 👥 CrewAI ⚙️ AutoGPT 🐍 Python 💚 Node.js ⚙️ VS Code 🌎 REST API

LLM cost to you

typed relations you control

Every

change tracked by agent

Real-time

agent-to-agent orchestration

NEW — No other tool offers this

Real-time agent-to-agent orchestration

LangGraph flags an issue. Cursor gets a webhook instantly. No polling. HMAC signed. Auto-disables on failures.

Webhooks

SSE stream

HMAC signatures

The problem with every other tool

Your agent forgets. Ours remembers why.

FLAT MEMORY

HYPERSTACK

Text search returns documents. Graph traversal returns answers.

See it in action

Watch your agent traverse the graph

Your agent asks “Why Stripe?” — and follows the links to find who decided, why, and what depends on it.

agent: “Why are we using Stripe?”

This is what the Graph Explorer does. From $29/mo →

See it in action

One question. Three linked cards. Zero follow-ups.

Your agent pulls typed cards, follows the links between them, and responds in one shot.

YOU

“Set up the new payments service and get it reviewed”

⚡ Agent searching HyperStack... q="payments service setup review"

⚙️ PREFERENCES

preference

prefers-typescript

TypeScript strict + Zod

typescript strict zod

Always TypeScript strict mode. Zod for validation. No any types.

🔗team-alice-backend

→ TypeScript strict with Zod ✔

👤 PEOPLE

person

team-alice-backend

Alice — Backend Lead

alice backend review

Alice owns backend services. Review PRs within 24h. Prefers async comms.

🔗prefers-typescriptdeploy-process-prod

→ PR assigned to Alice ✔

🔄 WORKFLOWS

workflow

deploy-process-prod

Deploy: Production flow

deploy staging checklist

PR → staging auto-deploys → QA signs off in #releases → Merge to main → prod.

🔗team-alice-backend

→ Staging deploy triggered ✔

AGENT

“Created payments service in TypeScript strict with Zod validation. PR assigned to Alice for backend review. Staging auto-deploys — I’ll ping #releases for QA sign-off.”

3 cards loaded 3 cards linked ~195 tokens 0 follow-ups

How it works

Three steps. That’s the whole product.

①

Store

Agent learns something? Save it as a card.

②

Next session, agent searches for what’s relevant.

③

Remember

Agent gets exactly the context it needs. ~350 tokens, not 6,000.

Here’s what that looks like in code ↓

cURL

Python

JavaScript

# 1. Store a memory
curl -X POST "https://hyperstack-cloud.vercel.app/api/cards?workspace=default" \
  -H "X-API-Key: YOUR_KEY" -H "Content-Type: application/json" \
  -d '{"slug":"project-api","title":"API","stack":"projects",
       "keywords":["fastapi","python"],"body":"FastAPI 3.12 on AWS"}'

# 2. Search memories
curl "https://hyperstack-cloud.vercel.app/api/search?workspace=default&q=python" \
  -H "X-API-Key: YOUR_KEY"

# 3. Get all cards
curl "https://hyperstack-cloud.vercel.app/api/cards?workspace=default" \
  -H "X-API-Key: YOUR_KEY"

# pip install requests
import requests

API = "https://hyperstack-cloud.vercel.app"
KEY = "YOUR_KEY"
h   = {"X-API-Key": KEY, "Content-Type": "application/json"}

# Store
requests.post(f"{API}/api/cards?workspace=default", headers=h,
  json={"slug":"project-api", "title":"API",
        "stack":"projects", "keywords":["fastapi"],
        "body":"FastAPI 3.12 on AWS"})

# Search
r = requests.get(f"{API}/api/search?workspace=default&q=python", headers=h)
print(r.json())

// Node.js / Deno / Bun
const API = "https://hyperstack-cloud.vercel.app"
const KEY = "YOUR_KEY"

// Store
await fetch(`${API}/api/cards?workspace=default`, {
  method: "POST",
  headers: { "X-API-Key": KEY, "Content-Type": "application/json" },
  body: JSON.stringify({
    slug: "project-api", title: "API",
    stack: "projects", keywords: ["fastapi"],
    body: "FastAPI 3.12 on AWS"
  })
})

// Search
const r = await fetch(`${API}/api/search?workspace=default&q=python`,
  { headers: { "X-API-Key": KEY } })

The math

Real savings, real math

Based on Claude Sonnet at $3/M input tokens, 100 agent messages/day.

Without HyperStack

~6,000

tokens every message

6,000 tok × 100 msgs/day
= 600K tokens/day
× 30 days = 18M tokens/mo
× $3/M = $54/mo input alone
+ output tokens & re-explanations ≈ $270/mo

→

With HyperStack

~350

tokens per message

350 tok × 100 msgs/day
= 35K tokens/day
× 30 days = 1.05M tokens/mo
× $3/M = $3.15/mo input
+ $29 Pro plan = $32/mo total — 96% less

Pricing based on Claude Sonnet 4 ($3/M input). GPT-4o and Opus savings are even higher.

Why HyperStack

Built different. Built better.

Other memory tools charge per operation, lock you into one platform, and need their own LLM calls. We don’t.

	HyperStack	Mem0 / Zep
Semantic + keyword search	✔ Hybrid (pgvector)	Semantic only
LLM cost per memory op	$0 to you	~$0.002/op on your API key
Auto-capture & auto-recall	✔ Built-in	Varies
Works with	MCP + LangGraph + REST	Usually 1 platform
Team sharing	✔ Up to 20 members	Rare / enterprise only
Structured cards	✔ Slugs, stacks, keywords, types	Unstructured blobs
Card types	✔ person, project, decision, event…	No typing
Card links & graph	✔ Link cards, traverse connections	Flat storage, no relationships
PII / secrets redaction	✔ Built-in rules	Not enforced
Setup time	30 seconds	2–10 minutes
Free tier	10 cards forever	Limited or paid only
Pro price	$29/mo — graph included	$99+/mo usage-based
Funding required	$0 — self-funded	$23M+ in VC

Integrations

Drop into your stack in 30 seconds

MCP server, LangGraph integration, or raw REST API. Pick your path.

🔌

MCP Server

✓ Tested in Cursor

npm install -g hyperstack-mcp

Works with Cursor, Claude Desktop, VS Code, Windsurf. Add 5 lines of JSON config and your agent gets persistent memory.

npm v1.1.0

🤖

LangGraph

✓ 3 lines to add memory

pip install hyperstack-langgraph

Pre-built tools + ReAct agent. Zero dependencies. Your LangGraph agent gets search, store, graph traversal, and time-travel.

PyPI v1.0.0

🌎

REST API

✓ Any language, any agent

pip install hyperstack-py

Python SDK or raw HTTP. cURL, fetch, requests — if it speaks HTTP, it works. No framework lock-in.

PyPI REST

Also on: npm · PyPI · ClawHub

Ready?

10 cards free. Graph from $29/mo. Start free, upgrade when ready.

What 10 free cards gets you

2-3 team members

2-3 decisions

2-3 preferences

1-2 projects

Enough to try it. $29/mo unlocks graph + 100 cards →

← Home

Documentation

📖 HyperStack Docs

Everything you need to give your AI agent persistent memory.

CORE CONCEPT

🃏 What are Cards?

Cards are small, structured chunks of knowledge your agent stores and retrieves. Think of them as index cards for an AI brain. Each card has:

slug

Unique ID — same slug = update, not duplicate

title

Human-readable name for the card

keywords

Tags for exact-match search

stack

Category folder (projects, people, etc)

cardType

person, project, decision, preference, workflow, event, account

links

Connect cards by slug — build a knowledge graph

📁 Stacks Categories for your cards

📦 projects

Tech stacks, repos, architecture

👤 people

Teammates, contacts, roles

⚖️ decisions

Why you chose X over Y

⚙️ preferences

Editor, tools, coding style

🔄 workflows

Deploy steps, review processes

📄 general

Everything else

🔗 Context Graph NEW

Cards aren’t flat. They connect. Type your cards, link them together, and traverse the graph.

Card Types

person, project, decision, preference, workflow, event, account. Filter cards by type.

Card Links

Connect cards by slug. “decision-auth” links to “team-alice” and “project-webapp”.

Link Resolution

GET a card → linked cards resolved automatically with metadata.

Graph Traversal

GET /api/graph?from=slug&depth=2 — follow connections up to 3 hops.

// Create a typed card with links
POST /api/cards?workspace=default
{
  "cardType": "decision",
  "links": ["team-alice", "project-webapp"],
  "slug": "decision-auth-clerk",
  ...
}

API REFERENCE

POST /api/cards?workspace=default

Create or update a card. Body: {slug, title, body, stack, keywords[]}

GET /api/cards?workspace=default

List all cards in a workspace

GET /api/search?workspace=default&q=term

Hybrid search: keywords + semantic matching (pgvector)

DEL /api/cards?workspace=default&id=slug

Delete a card by slug

All endpoints require header: X-API-Key: your_key

🔌 Using with MCP (Cursor / Claude Desktop / VS Code)

Install globally and add config. Your coding agent gets persistent knowledge graph memory.

npm install -g hyperstack-mcp

Add to .cursor/mcp.json or claude_desktop_config.json

Set HYPERSTACK_API_KEY in the config env

🤖 Using with LangGraph

3 lines to add knowledge graph memory to any LangGraph agent.

pip install hyperstack-langgraph

from hyperstack_langgraph import create_memory_agent

agent = create_memory_agent(your_llm)

🐾 Using with OpenClaw

HyperStack works as an OpenClaw skill. Your agent reads the SKILL.md file and automatically knows how to create, search, and update cards.

Place SKILL.md in skills/hyperstack/SKILL.md

Set env: HYPERSTACK_API_KEY=your_key

Done — your agent reads the skill and handles everything

❓ FAQ

How does my agent use HyperStack?▾

Your agent calls the REST API to store and search memories. Cards can be typed (person, decision, project) and linked to other cards by slug. Before responding, the agent searches for relevant cards. When it learns something new, it creates or updates a card. It can also traverse linked cards via the graph API to follow connections between knowledge.

What counts as a “card”?▾

Any piece of knowledge — a person’s role, a project’s tech stack, a decision rationale, your coding preferences. Cards are small (a few lines) and focused on one topic. Each card can have a type (person, decision, project, etc) and link to other cards, so your agent builds a connected knowledge graph over time.

How is this different from other memory tools?▾

Mem0 and Zep now have knowledge graphs, but they auto-extract entities using LLM calls (~$0.002/op, risk of hallucinated relationships). HyperStack lets you explicitly define cards and links — precise, zero LLM cost, instant. Plus time-travel debugging: reconstruct the entire graph at any timestamp to see what your agent knew when it made a bad decision.

I get “Unauthorized” when calling the API▾

Make sure you’re sending the header X-API-Key: your_key (not Authorization). Your key starts with hs_ and is in the API Key tab of your dashboard.

What fields does a card need?▾

Required: slug, title, body. Optional: stack (category), keywords (tags), cardType (person, project, decision, etc), links (array of card slugs to connect). If you POST the same slug again, it upserts. Use the graph API to traverse linked cards.

How does search work?▾

Hybrid search: keyword matching (instant, $0) + semantic search via pgvector (understands meaning). Ask “what auth do we use?” and it finds a card about Clerk even if “auth” isn’t a keyword — because it understands the meaning.

Can I use any AI agent?▾

Yes. Use the MCP server for Cursor/Claude Desktop/VS Code (npm install -g hyperstack-mcp), the LangGraph integration (pip install hyperstack-langgraph) for Python agents, or the raw REST API from any language. Works with Claude Code, LangChain, CrewAI, custom agents, or a simple script.

Is my data safe?▾

Your cards are stored in a secure PostgreSQL database with pgvector. Only your API key can access your data. We don’t read, sell, or train on your cards. HyperStack is self-funded — no VC pressure to monetize your data.

What happens at 10 cards on Free?▾

You can still read and search all your cards, but you can’t create new ones until you delete some or upgrade to Pro ($29/mo for 100 cards + graph traversal + visual explorer). The graph API and visual explorer are only available on paid plans.

PowerShell: curl doesn’t work!▾

PowerShell aliases curl to Invoke-WebRequest. Use Invoke-RestMethod instead, or run curl commands in Git Bash, WSL, or CMD.

Need help? Join our Discord

Portable memory with
a knowledge graph you control.

Real-time agent-to-agent orchestration

Your agent forgets. Ours remembers why.

Watch your agent traverse the graph

One question. Three linked cards. Zero follow-ups.

Three steps. That’s the whole product.

Real savings, real math

Without HyperStack

With HyperStack

Built different. Built better.

Drop into your stack in 30 seconds

MCP Server

LangGraph

REST API

Ready?

Reset password

Set new password

Start free. Unlock the graph.

Free

Pro

Team

Business

Activating your plan...

You're on Pro!

Payment received!

📖 HyperStack Docs

🃏 What are Cards?

Portable memory witha knowledge graph you control.

Real-time agent-to-agent orchestration

Your agent forgets. Ours remembers why.

Watch your agent traverse the graph

One question. Three linked cards. Zero follow-ups.

Three steps. That’s the whole product.

Real savings, real math

Without HyperStack

With HyperStack

Built different. Built better.

Drop into your stack in 30 seconds

MCP Server

LangGraph

REST API

Ready?

Create your account

Sign in

Reset password

Set new password

Start free. Unlock the graph.

Free

Pro

Team

Business

Activating your plan...

You're on Pro!

Payment received!

📖 HyperStack Docs

🃏 What are Cards?

Portable memory with
a knowledge graph you control.