---
title: "past.dev vs Graphon"
description: "Compare Graphon and past.dev by what each stores, the input it accepts, access control, what a read returns, and published BEAM results."
canonical: https://past.dev/vs/graphon
last-updated: 2026-11-13
---
# past.dev vs Graphon

Source: https://past.dev/vs/graphon

Graphon is file storage with search built in: you store video, audio, images and documents, and it answers questions about them with citations. past.dev is a memory API for agents: you send timestamped text, and recall returns ranked documents with dated source excerpts for the identity that asks. On BEAM, the long-term memory benchmark, past.dev scores higher at each size both have published, and its run has a public harness. Graphon accepts video and audio files. past.dev accepts text.

## What Graphon is

[Graphon](https://www.graphonlabs.ai/about) describes itself as a persistent intelligence platform for multimodal data that reads video, images, audio and documents. [Its documentation](https://docs.graphon.ai) presents the product as blob storage that replaces an object store such as S3. You put files in buckets, Graphon indexes each file, and you search the files or ask a question about them.

It has three read operations. Search returns spans of the stored files. Answer writes an answer with a citation for each source it used. Research runs longer, multi-step questions. There is a command line tool and an [MCP server](https://docs.graphon.ai/mcp). The about page lists three deployments: Graphon cloud, your own VPC, or on device.

## What past.dev is

past.dev runs as a managed service or, on an Enterprise agreement, in your own cloud. You `POST` text with the timestamp it happened: an email, a meeting transcript, a ticket, a chat message. `/recall` returns ranked documents with occurrence dates and source excerpts for the requested identity.

A stable source id replaces a source when it changes, and an optional recall timestamp reads memory as it was on an earlier date. past.dev accepts text only. For a call or a video, you send the transcript.

## Architecture differences

|  | Graphon | past.dev |
| --- | --- | --- |
| What it is | Blob storage with search and question answering | Memory API for agents |
| What you send | Files: video, audio, images, PDFs and other documents | Text plus the timestamp it happened, backdatable |
| Audio and video | Indexed from the file | Send the transcript as text |
| How you run it | Graphon cloud, your own VPC, or on device | Managed API, or your own cloud on an Enterprise agreement |
| Reading | Search returns spans; Answer writes a cited answer | `/recall` returns ranked documents with dated source excerpts |
| A source changed | Write a new version of the object | Reuse its stable source id for the new revision |
| Who can read what | Your application sends the reader's user and group names | past.dev keeps identities and audiences; a recall names the identity |
| BEAM run | Answers and grades published, with no harness, answer model or settings | Harness and per-question results public |

## Access control

In Graphon's storage API, an API key covers every bucket in its workspace, and the [tenant isolation guide](https://docs.graphon.ai/guides/tenant-isolation) has your backend decide which tenant a request can touch. [Graphon Engine](https://docs.graphon.ai/engine/concepts/access) has access rules for a folder or a file. A collection starts with access control off. When it is on, your application sends the reader's user and group names with each request, and Engine does not authenticate those names or expand group membership.

In past.dev, [identities and audiences](/docs/memory-api/access-control) are part of the API. You send a source for an audience, past.dev keeps who belongs to it, and a recall for an identity returns what that identity's audiences can read.

## How to choose

Choose based on the requirements below.

- **Choose Graphon** if your sources are video, audio or image files, or if you want one store that replaces S3 and answers questions about the files in it.
- **Choose past.dev** if your agent has to remember emails, meetings, chats and tickets over months, if each person must get only what they are allowed to read, or if you need a benchmark result you can rerun.
- **Both** leave your agent, your model and your prompts where they are. [What is agent memory?](/what-is-agent-memory) covers what a memory service has to do.

## Published figures

| BEAM history size | past.dev | Graphon |
| --- | --- | --- |
| 100K | 92.08% | 83.3% |
| 1M | 90.65% | 80.7% |
| 10M | 85.03% | 73.0% |

Each figure is the system's own published result: ours on [benchmarks](/benchmarks), Graphon's at [graphonlabs.ai](https://www.graphonlabs.ai/#benchmarks). Graphon publishes no figure at the other sizes, so those sizes are left out rather than shown as a gap. Graphon has published no harness, answer model or settings for its run, so a reader cannot rerun it. The past.dev [harness and per-question results](https://github.com/pastdotdev/benchmarks/tree/main/beam) are public.

Benchmarks are the part of a decision that is easiest to publish. The requirements above are what differs in use.

## Related

- [What is agent memory?](https://past.dev/what-is-agent-memory)
- [past.dev vs Supermemory](https://past.dev/vs/supermemory)
- [past.dev vs mem0](https://past.dev/vs/mem0)
- [BEAM leaderboard](https://past.dev/benchmarks/beam)
- [Memory API overview](https://past.dev/docs/memory-api/overview)
- [How we measure memory](https://past.dev/benchmarks)