Long-term memory for your AI.

Remember people, not just text. Give your assistant, agent or robot a memory that recalls what people said, what changed, and what they like, and shows the exact memories behind every answer.

LongMemEval-S · all 500 questions 91.8% full chat history with every question
LoCoMo · all 1,540 questions 90.3% answered by gpt-6-luna, a low-cost model
Memory sent per question ~6,200 tokens, picked from ~115,000 of history

Results

Scores on LongMemEval and LoCoMo

We ran every question in both benchmarks, once, and publish every answer.

LongMemEval-S

500 questions

Answered by gpt-6-luna. Graded by gpt-6.1-sol with LongMemEval's grading prompts.

Each question comes with about 50 past chats with an assistant, around 115,000 tokens. Ormem has to find the right ones itself.

459 of 500 right. Ormem sends about 6,200 tokens of memory per question.

By kind of question

  • All questions459 of 50091.8%
  • What the assistant said earlier55 of 5698.2%
  • What the person said68 of 7097.1%
  • Noticing when things changed75 of 7896.2%
  • Dates and timing121 of 13391.0%
  • Likes and preferences26 of 3086.7%
  • Combining several chats114 of 13385.7%

LoCoMo

1,540 questions

Answered by gpt-6-luna. Graded by gpt-6.1-sol with the same prompts.

Ten very long conversations between two people, each spread over many sessions.

1,390 of 1,540 right. Ormem sends about 5,600 tokens of memory per question.

By kind of question

  • All questions1,390 of 1,54090.3%
  • A fact from one session791 of 84194.1%
  • Combining several sessions253 of 28289.7%
  • Dates and timing285 of 32188.8%
  • Guessing what a person would do or think61 of 9663.5%

How we tested

  • The hard version of LongMemEval. Every question comes with the full chat history, not the "oracle" version that hands over only the chats with the answer.
  • Every question, one run. No answers combined from several runs, and no questions left out.
  • A low-cost answering model, gpt-6-luna. Grading by gpt-6.1-sol, since OpenAI is retiring gpt-4o, which many published scores use.
  • Tuned on 90 of the 500 LongMemEval questions. On the other 410, which we never tuned on, Ormem scored 91.2%.
  • Every answer is published: LongMemEval-S · LoCoMo. Each line has the question ID, Ormem's answer and the grade; the IDs match the official datasets.

Small context

Memory text sent to the AI per question (LongMemEval-S)

How it works

Two API calls: store and recall

Ormem searches the stored memories and sends the model only the relevant ones, 5,600–6,200 tokens per question.

Outer rings are newer memories

Question

Memories found

    Answer

    Example. The API returns the same sources as JSON.

    Send what happens

    Messages, notes and documents, as they arrive. Or import past chats from ChatGPT, Claude or WhatsApp in one go.

    Ask in plain language

    "What did Maya say about her mum's appointment?" Ormem finds the best passages, ranks them, and asks the model.

    Get an answer with sources

    Every answer lists the exact memories it used, with dates. When Ormem doesn't know, it says so instead of guessing.

    Store a memory
    # One key per customer, sent with every request
    curl $ORMEM_URL/v1/memories \
      -H "Authorization: Bearer $ORMEM_KEY" \
      -H "Content-Type: application/json" \
      -d '{
        "owner": "maya",
        "memories": [{
          "body": "Mum's cardiology check moved to Friday 3pm.",
          "speaker": "maya",
          "happened_at": "2026-10-02T09:14:00Z"
        }]
      }'
    Ask a question
    curl $ORMEM_URL/v1/recall \
      -H "Authorization: Bearer $ORMEM_KEY" \
      -H "Content-Type: application/json" \
      -d '{"owner": "maya", "question": "When is Mum'\''s heart check?"}'
    
    # Returns
    {
      "answer": "Friday at 3pm. It was moved there on 2 Oct.",
      "evidence_status": "supported",
      "citations": [{ "text": "Mum's cardiology check moved…",
                      "when": "2026-10-02T09:14:00Z" }]
    }

    A simple example

    How most AI remembers today

    Most AI apps remember by sending earlier messages, or a summary of them, along with every new one. The pile grows with each message, and summaries drop details. Ormem keeps everything and sends only what the new message needs.

    Sent to the AI

    The usual way

    Sent with the newest message

    With Ormem

    Sent with the newest message

    Features

    What it stores

    Conversations, in order

    Who said what, and when. The assistant's own past replies too, so "what was that recipe you gave me?" just works.

    Facts that change over time

    "Moved to Berlin" replaces "lives in Paris", and Ormem can still answer what was true back then.

    Likes and preferences

    Remembers what someone likes, dislikes and prefers, and uses it when they ask for a suggestion: "somewhere for dinner" leaves out the place they said they didn't enjoy.

    Documents with versions

    Care plans, policies, manuals. A newer version replaces the old one; if two sources disagree, both are shown.

    Bring past chats along

    Import exports from ChatGPT, Claude and WhatsApp. Re-importing the same file adds nothing twice.

    Hide, fade and forget

    Hide a memory so it only surfaces when its owner asks. Let old small talk fade to a summary. Delete permanently on request.

    Built with the API · Orson

    Turn your old Android phone into your personal memory device.

    Talk to it, and it remembers. Say "Orson", tell it something, and ask about it weeks later in your own words. One converse call per spoken turn does the rest.

    • Speech stays on the phone. Wake word, speech recognition and voice all run on the device; only text reaches Ormem.
    • Set up by Claude Code. Plug in the phone and say "set up my old phone as Orson".
    • Open source, so you can see exactly how a voice device uses the API.

    Trust

    Security and data

    • Every answer shows its sources with dates, and says when it can't find an answer.
    • Each person's memories are kept in their own store. One customer's key can never reach another's.
    • Memory text is never written to logs, and text inside a memory can't give the AI instructions.
    • Export or erase everything about a person at any time.
    • Answers from OpenAI's gpt-6-luna, a fast, low-cost model.
    • Private deployment for larger teams with strict data rules. Talk to us.

    The Ormem API is a separate product for developers building their own apps. The Ormem iPhone app is unchanged: your notes stay on your iPhone and never touch an Ormem server.

    The fine print

    • Each score is one full run. Repeat runs of the same setup usually land within about a point.
    • Scores from other memory products use different answering models, graders or benchmark versions, so they aren't directly comparable with these.
    • Token counts are estimated (characters ÷ 4) and include only the memory text.
    Early access

    Give your AI a memory.

    The Ormem API is invite-only for now. Tell us what you're building and we'll email you an invite.