KeyAmnesia
Giving an AI agent broad autonomy to run commands often means the agent also needs access to passwords and API keys, typically read from a plaintext .env file the agent can see directly. key-amnesia removes that exposure. Secrets are stored in an encrypted vault, using SecretBox (NaCl-based) encryption with Argon2id key derivation. When an agent triggers a command that needs a secret, the value is injected only into the child process's environment, never into memory the agent's own process can read, and any command output is censored for exact secret matches before it reaches the agent. Unlocking the vault requires a master password typed into a separate window the agent cannot read or control.
The project acknowledges real limits to this model: a command could still leak an obfuscated form of a secret, secret names themselves remain visible for listing, and processes running as the same OS user share underlying privileges regardless of the vault. The tool raises the bar rather than closing every gap, so the agent effectively gets amnesia about the secret's actual value.
As of August 2026 the repository has 35 stars and 0 open issues.
This is placed in Assess. It addresses a specific, real risk in giving AI agents command-execution autonomy, and its architecture (vault plus child-process injection plus output censoring) is a sound approach to that risk. It has not yet been used first-person in a workflow, which is the gate for Trial.
