Skip to content

PRIVACY.md

Latest commit

 

History

History
83 lines (52 loc) · 2.79 KB

PRIVACY.md

File metadata and controls

83 lines (52 loc) · 2.79 KB

Privacy

agent-memory is designed as a local-first memory runtime for AI agents.

What is stored

Depending on the commands and integrations you use, the SQLite database may contain:

  • approved semantic facts
  • candidate facts awaiting review
  • procedures and steps
  • episodes / work history
  • source records and excerpts
  • scopes and provenance metadata
  • lifecycle states such as candidate, approved, disputed, and deprecated

Default database path:

~/.agent-memory/memory.db

What is sent to models

agent-memory itself renders memory context. The host harness decides where that context goes.

For example, the Hermes pre-LLM hook returns bounded context to Hermes, and Hermes injects it into the current prompt. Codex/Claude prompt commands print prompt text that a wrapper can prepend to a model request.

By default, retrieval uses approved memories only. Candidate, disputed, and deprecated memories are excluded from normal prompt context unless you explicitly ask for them, for example with --status all during forensic review.

What is not sent by default

The core CLI does not intentionally upload your SQLite database to an agent-memory cloud service. There is no default hosted sync service.

agent-memory does not control what a downstream agent/model provider stores after your harness sends prompt context. Review the privacy policy of Hermes, Codex, Claude, or any other host you connect to.

Scope and path privacy

user:default is the recommended durable cross-project scope.

When the Hermes hook needs a project-local default and no explicit --preferred-scope is set, it can derive a cwd:<hash> scope from the runtime working directory. This avoids embedding raw local usernames or repository paths in prompt context.

Sensitive data guidance

Do not store secrets as memories.

Avoid approving memories that contain:

  • API keys
  • access tokens
  • passwords
  • private keys
  • customer data
  • production connection strings
  • confidential incident details that should never enter prompts

If sensitive data is accidentally stored, remove or rotate it like any other local secret exposure. If a secret may have been sent to a hosted model through prompt context, rotate it with the relevant upstream service.

Inspecting and deleting local data

Inspect whether the default DB exists:

ls -lh ~/.agent-memory/memory.db

Delete the default DB:

rm ~/.agent-memory/memory.db

This is destructive. Back up the file first if you may need its contents later.

Hermes config changes

agent-memory bootstrap may create or update:

~/.hermes/config.yaml

When modifying an existing config, the installer backs up changed files to *.agent-memory.bak. To disable the integration, remove the agent-memory hermes-pre-llm-hook ... entry from hooks.pre_llm_call.