Skip to content
How it works

It's just three steps: connect, it learns, every agent recalls.

Plug ULTRAMEMORY into your agents once. It quietly remembers what matters — and every agent on your team can recall it in milliseconds, all in sync.

The three steps[2 / 6]

Connect, it learns, every agent recalls.

The whole story in three plain steps — no vectors, schemas, or setup pain to think about.

  1. Step 01 · Connect

    Drop it in once.

    Add ULTRAMEMORY to any agent or tool — as an ambient layer, a 2-tool MCP plugin, or our SDK/REST. Your subagents inherit it automatically, so you wire it up once for the whole team.

  2. Step 02 · It learns

    It decides what's worth remembering.

    As your agents work, ULTRAMEMORY turns conversations into short, accurate facts — never dumping raw transcripts. It catches contradictions and keeps only what's true now, so your AI stays sharp instead of getting bloated.

    What it keeps — and what it drops
    1. Hundreds of back-and-forth messages — most of it is noise.

      messages: 184
  3. Step 03 · Every agent recalls

    Everyone's on the same page.

    Any agent on your team asks in plain language and gets back exactly the right memories in milliseconds — already trimmed to fit the model. Private notes stay private; shared truth is shared. They never fall out of sync.

    Recallsub-200ms
Every agent stays in sync[3 / 6]

One memory the whole team shares — private when it should be, shared when it matters.

This is where single-agent becomes multi-agent: a private note can be promoted into governed shared truth, and every agent keeps its own identity so you always know who remembered what.

Promotion gate
Private notepromotion gate Shared truth

quarantinedPoisoned or untrusted items bounce off the gate — they never reach shared memory.

Every agent has its own identity
shared bucketTeam leadResearcherWriterNew subagentReviewerWeb browseOld token
  • parent
  • subagent
  • peer
  • tool

Every agent has its own identity, so you always know who remembered what.

Add a new subagent and it's instantly caught up. Nothing is re-learned; nothing leaks.

Fast and accurate[4 / 6]

Why this keeps your AI fast and accurate.

Two must-wins, in plain language.

Fast

Recall lands in under 200ms, so your agents never wait on memory.

sub-200ms p95

Sharp

We hand the model a few clean facts, never a wall of old chat — so it stays accurate and never slows down.

Recall p95sub-200ms
ILLUSTRATIVE — OUR OWN NUMBERS
  • Recall p50fast path
  • Recall p95under 200ms
  • Accuracy (our eval)clean facts in

Dig into the numbers: instant recall · keeps your AI sharp.

Pick your altitude[5 / 6]

Pick how deep you go.

Minutes for a buyer, full control for a developer — your call.

01

Ambient / proxy

zero tools; it just works behind your agent.

02

MCP-minimal

two tools: remember + recall.

03

SDK / REST

full control in TypeScript or Python.

Works with
  • Claude Code
  • Codex
  • Cursor
  • MCP
  • Cline

See everything it plugs into →

For developers[6 / 6]

Want the 4-line version?

Buyer-first page, dev exit last — here's the real thing in a few lines: remember, then recall.

remember → recall
import { Ultramemory } from "@ultramemory/sdk";
const um = new Ultramemory();
await um.remember("Ships to the EU warehouse first.");
const facts = await um.recall("Where do we ship from?");
5-MIN QUICKSTART ↗Full developer docs →
Start free

Connect once. Every agent recalls.

Free to start, pay for what you use. Drop it into any agent or tool in minutes.