How to use the cognee-forget agent skill: remove data safely

What an agent skill is and what this one does

An agent skill is a folder containing a SKILL.md file that gives an AI agent instructions for one kind of task. The agent reads these instructions when a request matches the skill's description.

The cognee-forget skill covers deletion with forget(), cognee's single deletion API. The agent finds the right dataset, lists its documents, reads the stored content to confirm the match, and then removes only the requested item: one document, a whole dataset, or just the derived graph and vector memory while keeping the raw files.

The skill tells the agent to confirm with you before any deletion, because deletion cannot be undone.

Install the cognee-forget skill

Download the ZIP for this skill from this page and unzip it. The skill's only file is SKILL.md.

For a personal install, unzip it into ~/.claude/skills/. To scope it to one project, unzip it into that project's .claude/skills/ folder.

In the Claude apps you can instead upload the ZIP under Settings → Capabilities → Skills.

• Personal: ~/.claude/skills/
• Project: .claude/skills/ inside your project
• Claude apps: Settings → Capabilities → Skills

Set up what the skill needs

The skill assumes you already have cognee working. It needs one of the following installed and configured: the cognee Python SDK, the HTTP API, or cognee-cli.

The account or key the agent uses must have delete permission on the target dataset. For questions about access, the related skill cognee-permissions covers them.

• cognee Python SDK, HTTP API or cognee-cli installed and configured
• Delete permission on the target dataset
• cognee-permissions skill for access questions

Trigger the skill with natural requests

You do not need to name the skill. Ask for a deletion in plain language that describes the content, such as the topic of an outdated document or the name of a dataset. The agent will then list datasets and documents and read them before proposing what to delete.

Good requests name the content and the scope. Mention whether you want one document removed, a whole dataset emptied, or only the graph and vector memory cleared so it can be rebuilt later.

• Remove the document about our old 2023 pricing from the sales dataset.
• Forget everything in the onboarding dataset, the raw files can stay.
• Clear the graph memory for the support dataset so we can reprocess it, but keep the original documents.

Understand the deletion options

Choose the scope before the agent deletes. Deleting one document removes its graph nodes and edges that belong only to it, its vectors, and its data row; entities shared with another document remain until that document is also deleted. Deleting a whole dataset removes the dataset record, its data rows, and its graph and vector stores. Using memory_only keeps the raw files and data rows and resets the graph, vectors, sessions and pipeline status so the data can be re-processed.

To replace a document's content rather than delete it, the skill directs the agent to cognee.update() instead, since forget() is a delete.

• One document: forget(data_id=..., dataset_id=...)
• Whole dataset: forget(dataset="name") or forget(dataset_id=...)
• Rebuild graph later: forget(dataset="name", memory_only=True)
• Wipe all datasets you can delete: forget(everything=True), only on explicit request

Safety notes: scripts, network and permanent deletes

This skill runs no scripts and uses no network access of its own. It is a set of instructions that guides the agent's use of cognee, so the deletion itself happens through the cognee SDK, HTTP API or CLI you have configured.

Deletion is permanent. The cognee-cli forget command does not ask for confirmation, so the agent must confirm with you first. Be careful on shared multi-user deployments: forget(everything=True) removes data across every dataset the user can delete in the current tenant, including shared ones, and clears session caches for every user, not only you.

• Runs scripts: no
• Uses network: no
• CLI does not ask for confirmation
• everything=True affects shared datasets and all users' session caches

Tips for safe deletions

Ask the agent to show you the candidate documents with their names and short previews before any deletion. Match by meaning, not only by file name, since names are often generic hashes.

If you need a specific document removed, give the agent its exact id, and check afterwards that it no longer appears in the dataset listing. An unknown id can report success without deleting anything, so verification matters.

• Review the candidate list before confirming.
• Prefer exact ids for precise removals and verify with a new listing afterwards.
• Use cognee.update() for changed content instead of deleting.