Mnemonic

Privacy and Security

When we first began experimenting with personalized AI, we wrote a tool called memo to record our own agentic coding sessions. Memo records everything—agentic sessions, terminal commands, and snaps of the filesystem every N seconds. This is really unique data for us to experiment with, but we don't want so much data from you. Instead, we want the minimal amount of data we need to deliver personalized AI models: your agentic traces and access to your repo(s). For us to train a custom model that reduces how much back-and-forth you need with the models, we need a record of that back-and-forth. We get that via agentic traces. We need the underlying code these traces come from to properly contextualize them. Moreover, we ourselves are generally quite comfortable giving out agentic traces and repo access to startup tools that we try. We don't ask for any data that we ourselves wouldn't give.

We recognize sharing your code and work takes a lot of trust. We intend to earn it every day.

  1. If you ask us to delete your data (agentic traces, training artifacts, and trained models), we will do it right away. We only store traces in S3, including using S3 for backups. It's a one-click delete operation.
  2. Your data will will never be used to train a model for someone else. Your data is exclusively used to train models that you uniquely have access to.
  3. Our goal is to build a robust-enough training pipeline that no human ever needs to look at your data. While we are offering white-glove onboarding, we'll almost certainly need to look at your traces by hand to make sure we are training things properly.

We've written Mnemonic to capture data as follows:

  1. We only look for traces in ~/.claude/projects/ and ~/.codex/sessions/.
  2. Before uploading a session, Mnemonic checks which repository the work took place in. It uploads the session only if almost all of the relevant file paths point to a repository you authorized, and none point to an unauthorized repository. Simply mentioning a repository in a conversation does not count. Mnemonic also removes environment variables, content from sensitive files such as .env files and private keys, and recognized passwords, credentials, and tokens. These checks happen on your machine before anything is uploaded. The rules are built into the Mnemonic client.
  3. We upload data directly to a private, encrypted AWS S3 bucket using short-lived upload authorization issued by the Mnemonic server. This authorization permits writing only the expected upload and does not permit reading, listing, deleting, or writing elsewhere in the bucket.
  4. We don't download your repo from GitHub. As and when we need it, we'll pull it live. If you revoke access, we don't have a backup.

—Videh and Kalki, 2026-08-03