How to use the Forget Memories agent skill (forget)

What is an agent skill, and what does forget do?

An agent skill is a folder with a SKILL.md file that gives an AI agent step-by-step instructions for one task. When the skill is installed, the agent can follow those instructions when a request matches the skill's description.

The forget skill, from the rohitg00/agentmemory project under the Apache-2.0 licence, handles deletion of durable memories. It searches agentmemory for entries matching your request, shows you the memory IDs and titles with secret values redacted, and deletes only the IDs you confirm.

• It does not wipe whole sessions.
• It does not delete lessons. Lessons use a separate delete action.
• It does not handle observation IDs.

How do you install the forget skill?

Download the ZIP for the forget skill from this page and unzip it. Place the resulting skill folder in one of these locations, depending on how you want to use it.

For all your projects, put it in ~/.claude/skills/. For a single project, put it in that project's .claude/skills/ folder. In the Claude apps, you can instead upload the ZIP under Settings → Capabilities → Skills.

• Personal install: ~/.claude/skills/
• Project install: .claude/skills/ inside your project folder
• Claude apps: Settings → Capabilities → Skills, then upload the ZIP

What do you need before the skill works?

The skill depends on agentmemory being installed, with its memory_smart_search and memory_governance_delete tools available to the agent. Without those tools, the skill cannot run.

The skill's files reference a TROUBLESHOOTING.md guide for cases where those tools are missing. Check that guide first if the agent reports that a tool is unavailable.

• agentmemory installed and connected to your agent
• memory_smart_search tool available
• memory_governance_delete tool available

How do you trigger the forget skill?

Ask the agent in plain language. The skill's description lists phrases such as 'forget this', 'delete memory', 'remove that note', or a request to scrub specific data for privacy. Naming the thing you want removed helps the search find the right entries.

Good requests name a specific item, for example a note about an API key in a config file, or an old note from a named project. Vague requests like 'clear everything' will not be carried out as a full wipe.

• 'Forget that note where I pasted the API key.'
• 'Delete memory about the staging configuration in project app.'
• 'Remove that note about the old API key in config.'

What happens during a deletion?

The skill follows a fixed order so nothing is deleted by surprise. First it searches for matching memories. Then it lists what it found, redacting secret values. It asks for a clear yes or no. Only after an explicit yes does it call the delete tool, using the memory IDs you approved.

After the delete, the agent reports the actual number of deletions and any IDs it could not find or delete, reported separately.

• Search with memory_smart_search, limited to 20 results.
• Show memory IDs, titles and scope with secrets redacted.
• Wait for an explicit yes. Silence is not consent.
• Delete with memory_governance_delete using the confirmed IDs.
• Report deleted, not-found and failed IDs separately.

What does a deletion look like in practice?

In the worked example, the user asks the agent to forget a note where an API key was pasted. The agent searches with the query 'api key' in the project 'app', shows the memory ID and title such as 'Staging API configuration' without displaying the secret, and asks whether to delete it. After you confirm, it runs the delete with that ID and a reason of 'user privacy request'.

If the delete reports that one memory was removed, the agent reports one deletion. If it reports zero deleted and lists the ID as not found, the agent says the memory was not found.

What are the limits and safety notes?

Deletion is destructive and irreversible, so the skill is built around a preview and an explicit approval. Search results are relevance matches, not a full inventory, so some related data may not appear in the list. Check the list carefully before saying yes.

For whole-session erasure, the skill explains that this operation does not support it and does not pass a session ID. Lessons must be removed with a separate lessonId-based action. According to the skill's metadata, the skill runs no scripts and makes no network calls of its own.

• Destructive and irreversible: confirm each batch before approval.
• Search may miss some items; it is not a complete inventory.
• Whole sessions and all memory cannot be erased with this skill.
• Lessons need memory_lesson_delete instead.
• If a delete response is interrupted, recheck the IDs before retrying, since the write may have finished.

Tips for using the forget skill well

Use the 'recall' skill to find the exact memory ID before deleting when you are unsure what you want to remove. The 'remember' skill is the write side, and forget is its undo.

Be specific in your request and read the list before you approve it. Ask the agent to redact any secrets it shows you, and keep the reason field meaningful, because the skill records the reason you give it.

• Start with a precise search term rather than a broad one.
• Check the titles and scope before typing yes.
• Run a second search afterwards if you need to confirm the note is gone.