Skip to guide
PassControlby Vertias

Behest v0.1 beta

Memory custody: no stranger writes your agent's memory

Behest v0.1 beta. New memory waits at the border until you look at it.

Claude Code and Codex start every session by reading their memory files: CLAUDE.md, AGENTS.md, their rules and their auto memory. Whatever is written there, the agent treats as your instructions.

Now picture Tuesday. The agent reads a web page. The page has one line in it for the agent, not for you. The agent "helpfully" saves a note to memory. On Wednesday a fresh session starts, reads its memory, and follows that line. No prompt injection is visible in Wednesday's session. It was already inside.

Memory custody stops that note at the border.

Every write has an origin

The sidecar records every memory write, and what the session had read before it: a web page, an email, an MCP result, a cloned repository, an issue comment. A note written after outside content is an arrival, and arrivals wait at customs.

You choose what waiting means:

  • Hold: the next session's risky tools wait until you decide. Accept the note, or reject it.
  • Withhold: the session carries on, but the waiting note is cut out of what the model sees, replaced by a short marker. You review on your own schedule. This works in API mode. A subscription session is held instead.

The panel tells you where each arrival came from, for example "written by session … after reading a web page (mail.example.com)". Decisions need your dashboard sign-in with two-factor. There is no API for accepting.

The text never leaves your machine. The gateway keeps fingerprints and origins. passcontrol memory review shows you the words locally.

Attacked before it shipped

We built 34 poisoning attacks, crossing 30 ways to write memory with 8 ways a stranger's text arrives: a web page, an email, an MCP result, a file in a cloned repository, a commit message, an issue comment, a search result and a browser page. Every attack ran on a real disk, with real shell commands. The hook events that drive the sidecar were simulated, not taken from a live Claude Code session.

34 of 34 were held, still waiting a week later after a sidecar restart. The clean controls stayed clear.

In withhold mode, we checked what was actually forwarded to the model: 30 of 30 sessions carried no poison, while each arrival still waited for review. With the cutting switched off, all 30 failed, so the test measures the real thing.

Our own red-teaming, four internal rounds run with AI coding agents, plus the tests written while building it, found 13 ways around it. Each is fixed, with a test that stays in the suite. None found a path to "accepted" without the owner. The limits that remain are written down and pinned as tests, and the biggest is below.

What it costs you

Security that asks too often gets switched off. So we replayed a month of one developer's real work through the shipped rules: 16,028 tool calls in 18 sessions. It assumes the best case: an owner who decides at once.

  • Hold mode: 14 decisions in the month, a median of 3 in a full week.
  • Withhold mode, reviewing weekly: 5 decisions in the month. That is one a week, if every session runs in API mode. The price is that an honest note waits a median of four days before the model sees it.

That is one developer and one workflow. Yours will differ.

Where the edge is

Memory custody is a guard rail for agents with shell access, not a boundary. An agent that can run commands as you can also stop or change PassControl's own tools on your machine.

It covers memory files only. Vector stores and a provider's own memory are not covered.

Accepting an entry is your judgment that it is fine, not proof that it is true.

Try it

Memory custody is part of Behest v0.1 beta, on PassControl Cloud: passcontrol.vertias.eu. Turn it on in an agent's Memory custody panel, and run the sidecar.

Memory custody: no stranger writes your agent's memory · PassControl