AI agent memory that improves with time.
Bi-temporal knowledge graph, hybrid retrieval, sleep-time consolidation. The composite that wins each sub-problem in published memory research — packaged as a SaaS your agents can call in one line.
Sign up with email or Google and get an API key in under a minute.
What makes Memoria different.
Most memory products stop at “embed it and hope.” Memoria layers the three pieces that consistently win each sub-problem in published research, then runs them in a single sub-second pipeline.
Two clocks on every fact.
We track when a fact was true in the world and when Memoria learned it. Ask asOf: '2026-01-01' and get the world as your agent knew it then — full lineage on demand.
Dense + sparse + graph, fused.
Vector similarity, BM25 keyword search, and Personalized PageRank over the per-tenant graph — fused by reciprocal rank, then re-ranked by a cross-encoder. Sub-second p50 recall.
A pipeline, not a vector dump.
Seven extraction stages resolve entities, dedupe facts, detect contradictions, and assign event time — committed atomically. New writes can supersede old ones instead of silently piling up.
Drop into any agent loop.
Call remember() after each turn, recall() before the next. Use the SDK, plain REST, or the MCP server — same surface area, same guarantees.
- REST — signed with a
mem_live_*API key, scoped per environment. - MCP — one install line in Claude Code, twelve tools ready.
import { MemoriaClient } from '@premex/memoria';
const m = new MemoriaClient({ apiKey: 'mem_live_...' });
// write
await m.remember({
content: 'Stefan is a Senior Engineer at Premex.',
});
// recall — with optional time-travel
const facts = await m.recall({
query: 'Where does Stefan work?',
asOf: '2026-01-01',
});Built for the agents you're already shipping.
Memory is the bottleneck once an agent leaves the demo. A few shapes Memoria fits well today.
Agents that don't re-ask.
Account history, prior tickets, account-specific quirks — recall what you already know before you ask the user again.
Project conventions that stick.
Decisions, gotchas, and house style captured per-repo. The coding-pack MCP tools tag and retrieve them automatically.
An assistant that grows with you.
One workspace across sessions and devices. Time-travel queries make “what did I tell you last quarter?” a real answer, not a guess.
A single sub-second pipeline.
Writes run a seven-stage LLM cascade — entity extract, resolve, fact extract, dedupe, temporal, contradiction, commit — atomically per tenant. Reads parallel-fan-out dense + BM25 + PPR, fuse by RRF, then re-rank. Targets p50 under one second.
Start free. Scale when you're ready.
Quotas count successful calls per day. Paid plans are launching soon — the free tier is available today.
Hobby
Most popular- Everything in Free
- Higher daily quota
- Sleep-time consolidation
Team
- Everything in Hobby
- Multi-seat workspaces
- Priority support
Paid plans are launching soon. Until then, the Free tier is the only live plan — sign up and you'll land on the same dashboard either way.
Give your agent a memory.
Sign up with email or Google. Get an API key in under a minute. Start remembering.