Supported file types
RememberStack reads text. Every document is turned into Markdown before anything else happens, so what a deployment accepts depends on which formats it can turn into Markdown.
At a glance
On a self-hosted deployment:
| Format | As installed | With converters configured |
|---|---|---|
Markdown, plain text, subtitles (.md, .rst, .txt, .srt, .vtt, …) | Read; claims extracted | Same |
| Chat logs and transcripts, as Markdown or text | Read; claims extracted | Same |
| HTML, EPUB | Read; claims extracted | Same |
Word (.docx, .docm, .dotx), PowerPoint (.pptx, .pptm, .ppsx, .potx) | Read; claims extracted | Same |
Spreadsheets (.xlsx, .xlsm, .xltx, .xls), CSV/TSV, data files (Parquet, Arrow, SPSS, SAS, Stata, SQLite) | Profile: sheets, columns, counts and the first 5 rows; searchable, no claims | Same |
.ods | Profiled through LibreOffice (in the published image); stored and parked when the engine runs without it | Same |
.doc, .odt, .rtf, .ppt, .odp | Read through LibreOffice (in the published image); stored and parked when the engine runs without it | Same |
| PDF (digital, scanned or mixed) | Stored and parked without a Mistral OCR key | Every page is read with OCR when the key is configured; claims extracted |
Email (.eml) | Read; attachments listed, not read | Same |
Jupyter notebooks (.ipynb) | Markdown cells read (claims extracted); code cells searchable, no claims; outputs dropped | Same |
Code, configuration, logs (.py, .ts, .json, .yaml, .log, Dockerfile, …) | Read and searchable; no claims | Same |
| Text files of unknown kind | Read and searchable; no claims | Same |
| Images | File card (name, path, size, dimensions) | Read with image_ocr_description (PNG and JPEG; OCR plus a written description) |
| Audio and video | File card | Same |
Archives (.zip, .tar, .tar.gz, .7z, …) | File card; zip and tar list their members | Same |
| Other binary files | File card | Same |
| Web addresses (URLs) | No | No: download the page and send the HTML |
A Word or PowerPoint file over 100 MB, or a spreadsheet over 200 MB, gets a file card instead of a reading. A PDF over its 100 MB family limit fails conversion without a card (a PDF larger than the OCR provider accepts in one request, about 50 MB or 1,000 pages, is OCR'd in page batches); a spreadsheet over 50 MB is profiled from its sheet list and dimensions only. A text file over 1 MB is described by its size, line count, first 50 and last 20 lines instead of read in full.
The 100 MB and 200 MB limits protect the engine process, which reads the file itself. A deployment whose converter fetches files by a signed URL from S3 storage reads nothing in the engine, so it sends these files to that converter instead; only the converter's own input limit applies.
Formats and converters
A new self-hosted deployment reads the formats above locally, with no API
key, and describes images, media and archives with a card. Formats it
cannot read yet are accepted and stored, but they wait unprocessed. The
upload response tells you: its parked field is "no_route", and
remember ingest prints a warning.
To read PDFs, set REMEMBERSTACK_MISTRAL_OCR_API_KEY; every PDF page then uses OCR. To read images, map their type to a converter in
REMEMBERSTACK_SELFHOST_CONVERSION_ROUTES. File formats and
converters shows the setting, what each
converter does and costs, and how to release files that were stored before
you added a route.
The image description may quote a small number of clearly legible titles, names or labels as visual observations. OCR remains an independent text reading. Visual quotes do not replace or correct OCR, and image text is evidence rather than instructions. When readings conflict or a detail is unclear, inspect the original image. A new description-prompt version requires a fresh visual reading; completed OCR work can be reused.
Send the right type
A self-hosted deployment checks recognizable bytes before it uses the
extension or sent MIME as a format hint. A type that contradicts the bytes
is refused with HTTP 422. A text file with no recognizable family is
searchable but no claims are extracted from it. The remember CLI, Python
client and MCP ingest tool send a hint taken from the extension (the full
table). A file whose extension
does not say what it is needs the type passed explicitly:
client.ingest("notes/standup", mime="text/markdown")remember ingest notes/standup --mime text/markdownConversations
There is no separate conversation format. Write each conversation as one Markdown document with one line per turn, and send it like any other file. Ingest conversations and transcripts shows the layout.