How it works Oracle · built on Memoria Vs vendor memory Benchmarks Writing Get access
Theme

Persistent memory for AI

One memory for every AI tool you use.

Decisions, context and the reasons behind them carry across every session, every project and everyone on your team, in Claude, ChatGPT or whatever comes next. It stays yours when you switch.

Works with
  • Claude
  • ChatGPT Business & Enterprise
  • any MCP client

For developers

Already using Claude Code or Desktop?

Add Memoria as one MCP connector and every session shares the same memory: across machines, projects and your team's agents. Sign in once with OAuth; there's nothing to install per machine.

  1. 1 Get access. Memoria is set up per organisation, so we provision yours and send your connector URL.
  2. 2 Add it as a custom connector and sign in. Setup takes about two minutes.
Name:  Memoria
URL:   https://<your-org>.memoria.ashburn.studio/gateway/mcp

For everyone else

New to this? Here's the short version.

Memoria is memory your AI tools share. Tell Claude about a client on Tuesday, and on Friday ChatGPT already knows, and so does the colleague who started last week. It builds up as you work, so nobody has to maintain it, and it belongs to you rather than to whichever AI company you're using this year.

“Memoria is super helpful — remembering end of day tasks and stuff to pick up for tomorrow, and also checklists for the day.”

Ryan Rudkin, LoftWorks

Want an assistant in your team chat? Oracle is a chat assistant built on Memoria, for Slack and Telegram.

See Oracle

How it works

Memoria connects to your AI tools over MCP, the open standard that Claude and ChatGPT already support. Every tool reads and writes the same memory.

  1. 1

    Connect

    Add Memoria as a connector in each AI tool your team uses. Everyone signs in with their own account; there's nothing to install per machine.

  2. 2

    Capture

    As you work, decisions, context and the reasons behind them are saved to memory, from any tool and by anyone on the team.

  3. 3

    Recall

    Ask any connected tool and it answers from the same memory, with where each piece came from, so you can check it.

Sound familiar?

These aren't hypotheticals. They're Tuesday afternoon.

Your new hire asks Claude:

"Why did we go with flat-rate pricing instead of per-unit?"

Claude answers from Memoria: the original discussion from four months ago, the trade-offs you considered, and the client feedback that tipped the decision, with where each piece came from. Three seconds, not three meetings.

Institutional memory

You moved the team from ChatGPT to Claude last month.

"What did we agree with Harrison about the March deadline?"

The answer is still there. The memory lived in Memoria, not in either vendor's account, so changing tools didn't reset what your team knows.

Survives vendor change

A colleague picks up your project while you're on leave.

"Where did we land on the API rate limits?"

Their Claude session reads the same memory yours wrote to. No handover document, no waiting until you're back.

Shared across people

What Memoria does

One memory, shared and traceable. Two more capabilities are in development, and we're marking them that way.

Shared recall

Every connected AI tool and everyone on your team reads the same memory. Nobody starts from zero.

Provenance

Memories keep their source, so an answer can be traced back to where it came from.

Scoped sharing · Pensieve

Coming

Share selected memories with an approved outside party, such as a client or an auditor, without opening up the rest. Scope is set when a memory is written, never left to the reader.

Curated pages · Librarian

Coming

Turns accumulated memories into structured, wiki-style pages that stay current as the memory grows.

Where Memoria fits

Mem0, Zep and LangMem give an agent a memory. Memoria gives your team one, readable by every AI tool and every person, and owned by you.

It isn't a chat tool, a workflow tool or a CRM. It's the memory underneath them.

How it measures up

We ran Memoria through OmniMemEval, an open harness that tests memory systems under the same answer model and the same judge. Near the top on all three benchmarks, with no AI call when it stores a memory.

LoCoMo

accuracy

79.0

5th of 16

Leader MemOS 88.8 · Mem0 77.7

LongMemEval-S

accuracy

81.4

3rd of 14

Leader MemOS 89.2 · Mem0 56.0

BEAM-100K

nugget score

71.4

2nd of 13

Leader LCM-X 71.8 · Mem0 70.4

93%

of the memories our agents used in real work were written before that session started.

1 in 3

were more than a month old: knowledge that no single chat or context window holds.

Benchmarks: every system answers with gpt-4.1-mini and is graded by gpt-4o-mini; gaps of less than about 5 points are noise. Real work: 610 Memoria searches in our own Claude Code sessions, October to November 2026.

See the rankings and charts →

What's underneath

Built for teams that care how it works

If you're evaluating the stack, here's what matters. Memoria is a memory layer your agents reach over MCP, so the tools on top can change without the memory changing.

MCP-native

Any MCP-compatible agent gets organisational memory. Claude, GPT, Gemini, Llama — no vendor lock-in.

Semantic search

Recency-weighted retrieval surfaces what's relevant, not just what matches keywords. Ask why, not just what.

Knowledge graph

Decisions linked to rationale, trade-offs, and outcomes. Context compounds — it doesn't just accumulate.

Model-agnostic

Defaults to Anthropic for agentic and chat capabilities. Supports local models if you prefer to keep everything on-prem. Swap providers without losing context.

Self-hosted

Your infrastructure, your data

Deploy via Docker on your own servers. Full control over storage, embeddings, and access. For teams with strict data requirements or existing infrastructure.

Cloud-hosted

We run it, you use it

Same capabilities, no ops overhead. We handle the infrastructure so you can focus on work, not deployment.

Common questions

Which AI tools does it work with?

Any tool that supports MCP connectors, including Claude (web, Desktop and Code) and ChatGPT (Business, Enterprise or Edu, which support custom MCP connectors with write access). MCP is an open standard, so new tools that adopt it can read the same memory without any change on your side.

Do I need to know what MCP is?

No. If you'd rather not touch connector settings, we'll set Memoria up with you, and your team keeps using the AI tools they already have.

Who can see what's stored?

Your organisation's memory is its own. Each person connects with their own sign-in, and by default any write or other unsafe action needs their confirmation. Sharing selected memories with people outside your organisation is coming as a separate, scoped capability.

What happens if we change AI vendors?

Nothing is lost. The memory lives in Memoria, not inside any vendor's account, so a new tool connects to the same memory and picks up where the old one left off.

Is my data secure?

Cloud-hosted by default — we manage the infrastructure so you don't have to. For teams with strict data requirements, a self-hosted option is available where everything runs on your own servers. Either way, your data is yours.

Should I just use my AI vendor's built-in memory?

For a single tool and a single team, vendor memory often works fine — and we'd recommend it for that. Memoria is for organisations whose work spans multiple AI tools, multiple chat tools, and multiple years, where institutional knowledge can't live inside one vendor's walled garden. Read: Should you use vendor memory? →

Is there a version that lives in our team chat?

Yes. Oracle is a chat assistant built on Memoria, for Slack and Telegram. See Oracle →

Get access

Memoria is set up per organisation, so there's a short manual step before you connect. Tell us who you are and we'll set up your organisation's memory and send your connector URL.

* Required