← All posts

Blog · On agents · November 12, 2026

MemoryLake’s memory passport

MemoryLake is a hosted memory service you can point at ChatGPT, Claude, or Gemini. What it stores, which numbers are the vendor’s own, and what I would test before trusting it.

By Timothy O’Brien3 minute readFiled under On agents

MemoryLake is a hosted memory service. You load in documents, spreadsheets, chat history, audio, and video, and it hands the relevant parts to whichever assistant you are using: ChatGPT, Claude, Gemini, Qwen, or anything that can call an API. The company calls your store a Memory Passport, and the pitch is that it goes with you when you change models. I’ve read the site, the pricing page, and the integration list. I haven’t run it.

What it is

MemoryLake sorts what it keeps into six types: background, fact, event, dialogue, reflection, and skill. Background is the part you write yourself, and the system treats it as read-only. Facts are checked against each other and versioned. Skills are procedures you worked out once and want reused.

There are plugins for Claude Code and Codex, a built-in connection for OpenClaw and Hermes Agent, and MCP, a command line tool, and a REST API for everything else. Seven public data sets come already indexed, including PubMed and arXiv papers, SEC filings, and US patents.

Pricing is by token. The free plan is 300,000 tokens a month, which the pricing page puts at about 40 PDF pages or 1,200 retrieval calls. Pro is $19 a month and Premium is $199.

The claims

The homepage leads with a 94.03% score on LoCoMo, a benchmark for remembering long conversations, and calls it the highest score in its comparison. MemoryLake ran that test itself and links a GitHub repo for the run. Until someone outside the company reruns it, I’m treating it as a vendor number. The same goes for 99.8% recall, 91% fewer tokens, 1.5 million users, and 15,000 enterprises.

The security claims are specific. MemoryLake says encryption keys are split among three parties so that no single one, the company included, can read your memory. It says you can export everything in one click, and that a delete is permanent with no backups kept. It lists SOC 2 Type 2 and ISO 27001 and says the audit reports are available on request. I haven’t seen them.

What I would check

In Nobody owns the memory I asked for one place where you can see what a system carries forward, labeled by where it came from, with a delete that means delete. MemoryLake’s site makes a promise on each of those. Every memory is supposed to carry its source, its creation time, and a log of changes, and the history is described as working like git, with commits, diffs, and rollback. If that holds up in use, it is more than the consumer assistants offer.

The part I’d test is what happens when two memories disagree. MemoryLake flags the conflict and settles it by a rule you choose: the newest wins, a preferred source wins, the higher confidence wins, or you decide by hand. In the homepage example, ChatGPT has recorded that a user prefers dark mode and Claude has recorded light mode, and the newer entry wins.

That works for a display setting. It is a harder call when the disagreement is between something you said and a correction an assistant gave you. One of the studies I cited in that earlier post found that memory extraction sometimes kept the user’s mistake and lost the correction. Under a newest-wins rule, which one survives depends on which was said last. Before trusting it with anything that matters, I would tell it something wrong, correct it from a second assistant, and then read what the fact memory says and what the history shows.

MemoryLake is in the registry. In the book, chapter 8 covers the memory controls in ChatGPT and Claude, and chapter 10 covers agent memory you can correct.