# Introducing Sponge Sponge is a private library of articles and PDFs that your own AI agent can research from, and it keeps saved sources apart from the findings you review. By Hraness ยท Published 2026-09-24 Drafted with AI from the source code and reviewed by Claude Opus 5.5 (claude-opus-5-5) editorial review. Sponge is a private library for articles and PDFs that your own AI agent can research from. You save sources, read and annotate them, and ask an agent such as Codex or Claude Code to investigate them and write in a document through Sponge's API. You can ask it to cite the passages it uses, so a claim points back to the saved text. Research with an AI assistant in a chat window leaves little behind. Spend an afternoon asking about coral bleaching field studies, and the assistant finds a few good papers and quotes a few passages; the next day you have a chat log of confident sentences with no way to open the pages they came from. In Sponge, the papers are saved in your library, and the passages an agent quotes point back to them. ## Keeping sources, findings, and agent notes apart An agent doing research needs scratch space: which sources it has looked at, what it suspects, what it plans to check next. A reader needs claims that someone reviewed and that point to a source. Sponge keeps these in separate places, so one cannot turn into the other by accident. - **Your library** holds the saved original, the readable copy, and your highlights and notes, each kept distinct so a saved answer never stands in for the source. - **Reviewed knowledge** lives in Sponge's own product database. Proposals from your agent stay proposed until a person accepts, rejects, or returns them, and making anything public is a separate step after that. - **Agent notes** stay with your own agent. Sponge's retired hosted agent kept its working memory in a separate, short-lived store built on [Oh](https://oh.computer): only the document owner could see it, it expired within a day of the run starting, and Sponge's documentation says remembered notes are not evidence. When you come back to a question, the material you saved and the findings you reviewed carry over. An agent's scratch notes do not. ## Who Sponge is for Sponge suits someone who reads primary sources and wants an agent's help without losing track of where each claim came from: a researcher gathering field studies, a writer checking what a report says, anyone who has had to redo a search because the first assistant's answer had no trail. It is a poor fit if you want a hosted answer engine. Sponge's hosted source questions, overviews, and research runs are retired. You bring your own agent, and it works through Sponge's API and a downloadable research skill. When Sponge saves a page, it does not sign in to websites, render the page in a browser, read scanned images, or acquire audio and video. ## What you can do today Sign in at [sponge.computer](https://sponge.computer), save a public web page, a PDF with extractable text, or a supported text file up to 12 MiB, and read, highlight, and annotate the saved copy. If Sponge could save only part of a page, the reader marks it as partial. Documents keep version history, and each edit your agent makes creates a new version you can restore. Sponge's home page gives an example request: ask your agent to find coral bleaching field studies with measurements and save the ones worth reading. You can then give the agent selected passages and ask it to cite the ones it uses, check its claims against the saved text, and say where coverage is missing. ## Capturing sources locally with Ghostget Sponge's guides point to [Ghostget](https://ghostget.com) for capturing sources on your own machine, using only access you are authorized to use. Ghostget, not Sponge, handles any sign-in and permissions, and nothing about those accounts is sent to Sponge. The capture is an ordinary Ghostget command: ```sh ghostget capture https://example.org/study study \ --media none --evidence none --output /absolute/private/captures --json ``` Sponge's research skill describes an importer that turns that capture into a local evidence record. The importer runs from a checkout of Sponge's private source repository, outside the hosted app, and works offline: it never launches Ghostget, a browser, a model, or a network request. You choose the passages you want to cite, and the importer checks that the saved text matches the capture's recorded length and fingerprint and that every passage appears at the position you gave. The record carries the capture date and whether the capture was complete or partial. A partial capture stays marked as partial, and the importer refuses a capture that is too large instead of cutting it short. These checks show that the local files agree with each other. They do not prove that the capture is authentic or that a passage supports a claim. The record stays in your local research files and does not upload anything to Sponge. To put a passage into a Sponge document, use the library and document tools. ## Where Sponge is going The longer aim is a deep research system that improves by evaluating its own research runs. That system is a research project and is not part of Sponge today. It would build on the pieces Sponge has: a library that keeps sources intact, documents with history, and reviewed findings that point back to passages. ## Status and limits Status: Preview. Documents, version history, review, and publishing are live at sponge.computer. Saving, reading, highlights, notes, and search in the library are implemented; capture on the live service, journeys across accounts and devices, and the full signed-in review and publishing journey are being verified. Private knowledge search through the API has fixed per-request limits, and there is no public API search. Your library is private, and only a document version or reviewed research you choose to publish becomes public. The full list is on [Status and limits](https://sponge.computer/docs/status). ## Sources - [Status and limits](https://sponge.computer/docs/status), checked 2026-09-24 - [Inspect agent memory](https://sponge.computer/docs#agent-memory), checked 2026-09-24 - [Sponge research skill](https://sponge.computer/skills/sponge-research/SKILL.md), checked 2026-09-24