Get started

Research

How it works

Pricing

Get started

Get started

Research

How it works

Pricing

v1.1

Shared memory for agents and the people they serve

Shared memory for agents and the people they serve

A memory API that turns agent conversations into structured, versioned facts your agents recall across tools, sessions, and teammates.

Your AI tool has access to everything, but it still doesn't understand your business. With XTrace, it does.

Get API key

Read Docs

Memory Hub

···

XTrace Memory API

···

XTrace Vector Database

···

Memory Hub

···

XTrace Memory API

···

XTrace Vector Database

···

Memory Hub

XTrace Memory API

XTrace Memory API

XTrace Vector Database

XTrace Vector Database

Overview

What it is, who it’s for, and why it’s different.

The API

Structured, versioned memory

Turns agent conversations into semantic and procedural memory — recallable in a single call and you can integrate in 5 minutes

Built for

AI-native vertical startups

We're built for B2B teams shipping agents into a specific industry — not a general-purpose toolkit you have to bend to fit.

What’s different

A belief-revision engine

Newer facts revise older ones, with first-class retractions and lineage — not just nearest-neighbor search.

Shareable, not just per-user

Group memory

One memory layer for whole teams of agents — and the people working alongside them — not single-user assistants.

Benchmarks

The best —
with Believe Revision System

Read the research paper

LOCOMO — long-term conversational memory

Higher is better. Evaluated on the public LOCOMO suite (Snap Research). Competitor scores from published vendor papers & blogs.

XTrace Memory

92.3%

Mem0 (selective)

91.6%

Zep

75.1%

Letta (filesystem)

74.0%

OpenAI Memory

~52.9%

Notes: Mem0 score = their 2026 token-efficient pipeline (mem0.ai/research). Zep = Zep's own corrected number per their rebuttal (getzep.com). Letta = their published filesystem-only baseline. OpenAI

Memory is estimated from Mem0's comparative "26% relative lift over OpenAI" figure.

LOCOMO — long-term conversational memory

Higher is better. Evaluated on the public LOCOMO suite (Snap Research). Competitor scores from published vendor papers & blogs.

XTrace Memory

92.3%

Mem0 (selective)

91.6%

Zep

75.1%

Letta (filesystem)

74.0%

OpenAI Memory

~52.9%

Notes: Mem0 score = their 2026 token-efficient pipeline (mem0.ai/research). Zep = Zep's own corrected number per their rebuttal (getzep.com). Letta = their published filesystem-only baseline. OpenAI

Memory is estimated from Mem0's comparative "26% relative lift over OpenAI" figure.

Five-minute quickstart

Messages → searchable facts

Install the SDK, ingest a turn, and retrieve it back as structured memory — the TypeScript client handles extraction and polling for you.

import { MemoryClient } from "@xtraceai/memory";

const memory = new MemoryClient();

// Save what your agent learns
await memory.ingest({ messages, user_id: "alice" });

// Recall it anytime, in any tool
const facts = await memory.search({
  query: "what does the user like?",
  user_id: "alice",
});

Get API key

Read Docs

How it works

One write. One recall.
A belief engine in the middle.

Click a node on the timeline to see how memory builds up with XTrace