The L2 cache for AI agents
The fast, predictive memory tier between your model and cold storage — it keeps the right context warm before your agent asks, survives every session, tool and restart, and proves nothing was altered behind your back. One config block; your AI never starts from zero again.
Latency and hit-rate measured on our own production substrate — not a customer benchmark.
The memory hierarchy
Your vector database is main memory. Parametric Memory is the L2 cache in front of it — the fast, predictive, verifiable tier that keeps the right context warm before your agent asks. Verify a memory yourself →
Setup
No Docker. No self-hosting. No SDK. Sign up, claim your key from your dashboard, and paste one config block into the AI client you already use.
Sign up and pick a plan — your substrate provisions automatically.
Claim your API key in your dashboard — your config block is ready to copy.
Paste one MCP config block. Your AI has memory — permanently.
{
"mcpServers": {
"parametric-memory": {
"type": "streamable-http",
"url": "https://silver-flat-6czf.droplet-mcp.nz/mcp",
"headers": {
"Authorization": "Bearer mmpm_live_••••••"
}
}
}
}What your AI gets
The benefit first. The mathematics that make it true, underneath.
Every fact, committed and verifiable.
Each memory is written to a SHA-256 Merkle tree the moment it arrives. Tamper with one atom and the root hash changes — so your AI can verify its own memory without trusting the server.
RFC 6962 compact proofs · 37% fewer tokens (4,102 → 2,580)
The next memory is already loaded.
A Markov chain learns which memories tend to follow which, and pre-fetches the context before the query lands. Recent memories weigh more; old paths decay. A vector database can delete — only a parametric memory can forget, which is why relevance doesn't rot as the store grows.
64% predictive hit rate on our substrate
Recall in a fraction of a millisecond.
LevelDB sharded across four independent Merkle trees keeps recall sub-millisecond. Professional and Team run on a dedicated droplet — zero contention, no noisy neighbours; every tier stays an isolated substrate.
0.022ms p50 · 0.046ms p95, measured in production
The client your AI already speaks.
A Model Context Protocol server with 11 tools over streamable HTTP. Works with Claude, Claude Code, Cursor, Cline and any MCP-compatible client. No SDK to learn.
11 MCP tools · OAuth2 + Bearer
Built for who's actually using it
Increasingly, the thing driving your dashboard isn't a person — it's Claude-in-Chrome, Operator, Atlas. So every button here has a stable handle and a documented API equal, and every page ships a machine-readable action manifest. A human clicks it; an agent calls it. Both just work.
A UI control and an API endpoint that does exactly the same thing.
A published actions manifest and llms.txt so agents can find their way.
Semantic HTML and labels checked in CI — the build fails without them.
Trust, earned not asked
Drop a signed memory snapshot into your browser and verify it yourself — no API key, no account, not one line of our code in the loop. If a single atom was altered, the check fails in front of you. That's the whole point: you shouldn't have to trust us. Ask any memory vendor for the same artifact.
We run our own company on it — 1,000+ Merkle-sealed atoms in production, and growing.
Pricing
Every plan is an isolated substrate with its own Merkle tree and API key — Professional and Team run on dedicated infrastructure. Pick by how much your AI needs to remember.
Starter
$5/mo
1,000 memories
Solo
$9/mo
10,000 memories
Professional
$29/mo
100,000 memories
Team
$79/mo
500,000 memories
Enterprise & self-hosted
Dedicated clusters, on-prem deployment, custom capacity and SLAs.
All prices in USD · 7-day money-back guarantee · cancel anytime
Get started today
Start in minutes.
Never start from zero again.
Your own isolated substrate, cryptographic proofs and predictive recall — your own Merkle tree, not a row in someone else's database.
Get your instance — $5/mo7-day money-back guarantee · cancel anytime