v1.0 · public beta
Benchmarks
The numbers that matter in production.
Under the hood
Why not just a vector DB?
A vector DB answers "what looks like this?" A memory engine answers "what's true right now?" Those are different queries, and they need different data structures.
Dimension
Vector DB + RAG
"what looks like this?"
XMem
"what's true right now?"
✕
✓
✕
✓
✕
✓
✕
✓
Integration
✕
✓
Architecture
One write. One recall.
A belief engine in the middle.
Your agent sends raw events. XMem extracts facts, resolves entities, supersedes stale beliefs, and serves back a ranked, scoped, source-tagged context window.
your
agent
writes events
→
XMem engine
→
ranked context
injected into LLM
$ curl -X POST https://api.xmem.ai/v1/recall \
-H "Authorization: Bearer xm_live_..." \
-d '{"agent_id":"support-bot","user_id":"u_4713","query":"reach them?"}'
→ {
"facts": [{
"content": "prefers email over SMS",
"confidence": 0.94,
"source": "chat:2026-04-18",
"supersedes": ["fact_8a1c..."],
"ttl_days": 180
}],
"prompt": "User u_4713 prefers email over SMS (src: chat 2026-04-18).",
"tokens": 42, "latency_ms": 67
}
Core primitives
Six objects. One coherent mental model.
Everything in XMem is built from these. Learn them in an afternoon; never hit a dead end.
Built for
Every agent that holds a conversation longer than a goldfish.
Support agents
Know the user's plan, their last three tickets, and the policy changes since they signed up. Stop asking "what's your order number?"
Coding agents
Remember which patterns this repo favors, what conventions were deprecated last month, and which files the user owns.
Voice agents
<80ms p50 is the latency budget you need. Facts-only recall means short prompts — fewer tokens, lower cost, faster TTFB.
Research agents
Track what's been read, what's been ruled out, and which claims are still load- bearing. Supersede chains = reproducible research.
Companions
Users want an AI that actually knows them. Not one that forgot their dog's name between sessions.
Workflow / ops agents
Long-running agents that span days and tools. Lineage matters here more than anywhere — XMem gives you audit + replay.
For CTOs
The memory layer you'd build if you had a year.
Every team building agents writes some version of this themselves. Chat history truncation. De-duping facts. Hand-rolled entity resolution. Deciding whose contradiction to trust. It's six months of distraction from your actual product.
XMem is that layer — built, battle-tested, benchmarked, compliant. You ship the agent. We ship the brain.
HE
Homomorphically encrypted by default
Our own encrypted vector database keeps embeddings searchable while encrypted — we never see your data in the clear.
0
Zero training on your data. Ever.
Your facts are yours. Never used to train models. Export-on-exit as JSON.
SDK
Python SDK — open source.
Apache 2.0 licensed. Self-host the SDK; point it at our managed engine or your own deploy. TypeScript and Go coming soon.
⟶
Drop-in. No lock-in.
Three calls to integrate. One call to export everything. We don't hold your data hostage.
Pricing
Start free. Scale when it works.
Every plan includes every feature. You're only paying for volume.
Hacker
Side projects & hackathons.
$
0
free forever
10K adds / month
100K recalls / month
All primitives & SDKs
Community Discord support
Shared infra · p50 < 120ms
Startup
First agent in production.
$
49
/mo
+ usage beyond included
250K adds / month
2.5M recalls / month
Shared infra · p50 < 100ms
Email support · 1 business day
Private Slack channel
Usage analytics dashboard
Enterprise
Regulated, at scale.
Custom
talk to founders
Unlimited volume
VPC / single-tenant deploy
HIPAA · custom DPA · SSO
99.99% uptime SLA
Dedicated solutions engineer
Custom decay & policy models
Priced on
adds
(events written) and
recalls
(context requests). Everything else — facts, episodes, artifacts, supersede chains — is just how we think about what's inside.
Same engine, two surfaces
Your agent has memory. Your whole team should too.
Most XMem customers ship their agent, then roll memory out to the rest of the team. That's Memory Hub: the same belief engine, surfaced as a product humans can use.
you are here
XMem
Memory for your agent
A drop-in SDK. Your code, your agent, your stack. One engine per application.
·
Python SDK (TS + Go coming soon)
·
Fact-level belief revision
·
Scoped per agent · per user
·
$0 → $199 → custom