RememberStackremember.dev/docs

What is not built yet

This page lists what you might expect and will not find. Everything here is absent from the current release; nothing here has a date.

Ingest

  • Audio and video. RememberStack stores audio and video and describes each file with a card, but no converter turns speech or video into text.
  • Email attachments. An .eml file's attachments are listed by name and size; they are not read.
  • URLs. You cannot ingest a web page by its address; fetch the page yourself and upload the file.
  • Connectors and watched folders. Nothing pulls documents from another system or a folder automatically; the remember connectors commands and SDK connector methods call routes that no deployment serves, and the connectors-watched-directory install extra installs nothing.

Documents

  • Erasing a document. Deleting a document removes it from the memory but keeps its claims and stored original as history; the HTTP API, the SDK, the CLI and the MCP tools cannot erase a document's bytes.
  • Deleting one version. You can delete a whole document, not a single version of it.
  • Document filters on fact lookups and SQL functions. Claim and chunk search and the three context operations (claims_and_sources_context, facts_context, combined_context) take a documents filter, tags included; the direct fact lookups (/lookup/relations, /lookup/observations) and the SQL query functions do not. In SQL queries you can join on document_tags_live.
  • Metadata from other formats. Word, PowerPoint, PDF and email files give their title, people and dates, and .xlsx workbooks their title, author and dates; HTML, EPUB and notebooks give only a title, and other formats give none.
  • References found in the text. RememberStack does not yet find references itself (links, citations, "see section 4.2"); every reference document_references returns is one you supplied with PUT /documents/{doc_id}/versions/{version_id}/references.
  • Periods, references and section history from the CLI and MCP. The CLI declares periods only at remember ingest, and the MCP server only through ingest; neither can change periods later or supply references. The CLI has no section history or references command; the MCP server has the section_history and document_references tools.
  • Waiting for readiness from the CLI. No remember command waits for a document to be processed; use the SDK's wait_for_readiness or the MCP pipeline_readiness tool.

Knowledge

  • Compiled pages. RememberStack does not write summary pages that combine what it knows about a topic; the envelope's pages field and the compiled grain exist but stay empty, and the k install extra installs nothing.
  • Opening an exact source through a client. No API/SDK/MCP source_open operation returns a whole original or historical representation; agents read passages through search and adjacent_chunks (a self-hosted engine can publish its sources as files, see Filesystem views).

Clients

  • Async Python client. The remember package is synchronous only.
  • Other languages. A Node.js TypeScript/JavaScript client is available as repository source. There is no Go client; use the HTTP API.

Running the engine

  • Retained-storage measurement. The engine does not produce qualified consistent storage captures; its evidence-encoding helpers are not a storage meter.
  • Kubernetes. There is no Helm chart or Kubernetes manifest; the supported deployment is Docker Compose (Install with Docker Compose).