Memory infrastructure for AI

Give your AI a memory it can trust.

MemoryOS is the memory infrastructure for AI applications — storing what matters, updating what changed, and giving every retrieved memory a reason.

Build with MemoryOS See how memory works
p95 latency
11 ms
p95 latency
contradictions
0.0%
contradictions
PII leakage
0.0%
PII leakage
passing tests
97
passing tests

Act 02 — The problem

AI can generate.
It still doesn’t remember.

Chat models answer from their training data. Between sessions, they forget everything about your user — including what they were just told.

Without memory

forgets
I prefer tea now.
Got it — anything else?
3 weeks later
What should I drink?
I’m sorry, I don’t know what you drink.

Every session starts from zero. The AI can generate — but it can’t carry a preference across a single week.

With MemoryOS

remembers
I prefer tea now.
+ preference.drink → tea
3 weeks later
What should I drink?
Tea.
memory retrieved · source user_stated · confidence 0.94

The engine stored the preference, superseded the old value, and retrieved it with a reason — three weeks later.

That is memory.

Act 03 — See inside

Make memory visible.

Know what your AI remembers, why it remembers it, and what changed.

memoryos // event loglive
  2. 00.850INDEXmem_117
  6. 04.250CONFLICTcoffee ↔ tea
  9. 06.800AUDITPASS
  16. 12.750CONFLICTindie ↔ jazz
  19. 15.300COMPLETE18 memories
  20. 16.150AUDITPASS
ACTIVENEWSUPERSEDEDDELETEDREDACTEDCONFLICT

Act 04 — Explainability

See every choice it makes.

Every write and every retrieval can be traced to a decision — extracted slot, detected conflict, resolution rule, provenance, confidence. Nothing happens silently.

MemoryOS / SESSION 001
> ingest [input] "Actually I switched to tea now."
> extract [slot] preference.drink
> pii [scan] clean
> conflict [existing] coffee
> resolution [rule] coffee → superseded
> update [write] tea → active
> provenance [source] user_stated
> confidence [score] 0.94
memory updated

Act 05 — Developer experience

One layer.
Persistent memory.

MemoryOS is a layer, not a rewrite — extract, resolve, store, retrieve and explain, while your app keeps its existing stack.

APP
MemoryOS
Extract
Resolve
Store
Retrieve
Explain
LLM
app / memory.tstypescript
import { MemoryOS } from "@memoryos/sdk";

const memory = await memoryOS.remember({
  userId: "user_123",
  input: message
});

Illustrative — the public SDK ships with the API surface, which is on the roadmap. Today the engine runs as a Python package on Postgres + pgvector.

memory (result)
{
  "type": "preference",
  "key": "drink",
  "value": "tea",
  "confidence": 0.94,
  "source": "user_stated"
}

Act 06 — Trust

Your AI remembers.
But can you prove what it remembers?

Measured, not marketed

Numbers below come from the engine’s own acceptance benchmark — the D3 replay suite that ships with the repository.

11 ms
p95 retrieval
0.0%
contradictions
0.0%
PII leakage
97
passing tests

AUDITABLE · DETERMINISTIC · TRACEABLE · PRIVACY-AWARE

Act 07 — Where it fits

Memory belongs in the stack,
not in the prompt.

A memory layer between your model and your users — for anything where a forgotten preference is a real cost.

Assistants & copilots

Project context, decisions and preferences that survive sessions. Hybrid retrieval with provenance-weighted ranking — what the user stated always outranks what the assistant inferred.

Health · finance · family

Allergies, budgets, constraints. Correct facts win at write time, and PII is scrubbed to placeholders before it ever reaches the store.

Organizations

Retention and deletion accountability. Deletion is lineage-aware and physical, memory is per-tenant isolated, and every decision leaves a trace.

Platform operators

Deterministic admission and retrieval — p95 well under the 150 ms budget, zone-budgeted context injection, and no LLM calls on any write, read or delete path.

Not a chat-log store · not vector-only · not a full assistant — a memory layer

Your AI is getting smarter.
Give it a memory.

Build with MemoryOS

Postgres + pgvector core97 tests · all greenNo LLM on the write path