Remembrall

Remembrall has shut down

Shut down · 410 Gone

By Tyler Mayberry · 1 October 2026 · Not affiliated with Reducto or the original Remembrall

If a link brought you here from LangChain's docs or the Remembrall repository: the long-term memory API that lived at this address went offline in August 2024. Its dashboard, sign-in, API keys and stored memories went with it, and nothing on this site can sign you in.

What to do now

  1. Point your client back at your model provider's base URL.
  2. Remove the x-gp-* headers.
  3. Rotate any OpenAI key you ever sent through Remembrall.

Then pick a replacement from the options below.

Code still pointing at
https://remembrall.dev/api/openai/v1? Requests there now get an immediate 410 Gone error. They aren't forwarded to OpenAI, read or stored.

What Remembrall was

Remembrall described itself as "long-term memory for LLMs": add two lines to your OpenAI call, and replies would be personalised from a user's past conversations or your documents. It began as a prize-winning project at the Modal hackathon in New York and was published by Reducto (Y Combinator, Winter 2024), whose product today is document parsing.

It was a proxy in front of OpenAI's Chat Completions API. You changed the base URL to https://remembrall.dev/api/openai/v1, kept your own OpenAI key in the Authorization header, and added:

  • x-gp-api-key: your Remembrall key;
  • x-gp-remember: a user ID, usually an email, which turned on memory for that user;
  • x-gp-context: a document set uploaded in the dashboard, for retrieval;
  • x-gp-short: replayed the last 20 messages as short-term history.

After each call, a GPT-3.5 function call pulled facts out of the conversation and stored them as embeddings. Before the next call for the same user, the five closest memories were added to the prompt. There was a free tier, a $50-a-month Starter plan and, by 2024, an enterprise plan from $2,000 a month.

The Internet Archive last saw the site working on 13 August 2024. By 27 August a domain-parking page had replaced it. I found no public shutdown notice. LangChain's integration page still describes it, with links to the old sign-in and settings pages.

Moving off it

  • From LangChain: remove openai_api_base and the x-gp-* headers from ChatOpenAI, and use LangGraph's long-term memory. Keep memories in a Store namespaced by the same user ID you sent as x-gp-remember, save facts during or after each turn, and search the store before replying.
  • From a plain OpenAI call, in Python: the closest swap is Mem0's OpenAI-compatible client. It has the same chat.completions.create shape, with user_id in place of x-gp-remember, and it saves and recalls facts on each call. From other languages, use Mem0's REST API and add and search memories yourself.
  • Document contexts (x-gp-context): these were retrieval over your files, not memory. Any vector store and retriever does that job.

Ways to give an agent memory now

Checked 30 September 2026. All of these can run on your own hardware; the "where it lives" column is where your agent's memories end up.

OptionWhat it isWhere it lives
LangGraph Store (MIT)Long-term memory as JSON documents under namespaces you choose, such as a user ID, with semantic search.Postgres, MongoDB or Redis in production; in memory for development.
Mem0 (Apache-2.0)A memory layer: add and search memories by user, agent or run. An open-source library, a self-hosted server or a hosted platform.A local Qdrant and SQLite by default; Postgres with pgvector for the server; Mem0's cloud for the platform.
Letta, formerly MemGPT (Apache-2.0)A platform for stateful agents whose memory is a git-backed set of files the agent edits itself.A local server, or Letta Cloud.
Graphiti (Apache-2.0)A graph of people, things and facts that tracks when each fact was true. Zep is the hosted service built on it; Zep's own open-source edition is deprecated.Your Neo4j, FalkorDB or Amazon Neptune.
MCP memory serverThe official reference server: a small knowledge graph any MCP-capable agent can read and write.One JSONL file.
OpenClaw (MIT)A self-hosted agent that keeps memory as plain Markdown it writes and searches, with "no hidden state".MEMORY.md, USER.md and daily notes in its workspace folder.
Hermes Agent (MIT)A self-hosted agent with a small, curated memory it edits itself, plus searchable history of past sessions.~/.hermes/memories/ and a local SQLite file.

If you're building an app, Remembrall's job now falls to the first four. If what you wanted was one assistant that remembers you, the last two do it with memory you can open and read in a text editor, because it's ordinary files on the computer the agent runs on.

Memory on a computer you own

A self-hosted agent keeps its memory wherever it runs. Put it on a spare mini PC or old laptop and its notes, history and logins stay in your house, not in someone else's database, which is the failure Remembrall's users ran into. Self-Hosted AI Agent helps you check whether a computer you already have is enough, and Hermes vs OpenClaw compares the two agents above, including how each one remembers.

I set up and support agents on computers people own through my business, Cirlet.

Questions

Can I get my old Remembrall memories or logs back?

Not from here. The service and its database went offline in August 2024, and the domain lapsed after that. Registering the domain doesn't come with any of the old data, accounts or keys.

Is this site Remembrall or Reducto?

No. Tyler Mayberry registered remembrall.dev on 1 October 2026, after it had lapsed. This page isn't affiliated with Reducto, Y Combinator or the people who built Remembrall.

Is the Remembrall code still available?

Yes. The source is on GitHub at reductoai/remembrall under the AGPL-3.0 licence. It has had no commits since October 2023 and depends on Supabase, Stripe and gpt-3.5-turbo-0613, a model OpenAI has since retired, so it's a reference rather than something to run.

Should I rotate an OpenAI key I used with Remembrall?

Yes, if you haven't already. Remembrall worked as a proxy, so every request carried your OpenAI key through it. Nothing here reads or keeps requests, but rotating a key that passed through a service that no longer exists is cheap and sensible.

Sources