lobotomize

Lobotomize

Prune the agent's long-term memory. Show every memory with its age and purpose, let the user pick which to forget, then hard-delete the chosen ones and clean up after them.

Step 1: Locate the memory store

Find the memory directory for the current project — the folder holding one-fact-per-file memories and a MEMORY.md index. Enumerate every memory file in it, excluding MEMORY.md itself.

• No memory directory, or it holds only MEMORY.md (or nothing) → tell the user there's nothing to prune and stop.

Done when: you have the path to the store and the list of memory files.

Step 2: Gather each memory's facts

Read every memory file and the MEMORY.md index. For each memory collect:

• Name — the frontmatter name: slug (fall back to the filename).
• Age — the file's last-modified timestamp expressed relative to today (e.g. 3d, 2w, 5mo).
• Why it exists — a one-line gloss of its purpose, drawn from the frontmatter description:, the MEMORY.md hook, and the body (type, the **Why:** line for feedback/project memories).

Do not guess a purpose you can't support from the file — say unclear and let that be a signal for purging.

Done when: every memory has a name, an age, and a purpose line.

Step 3: Render the table

Delegate to /tabular-analysis to present the comparison. Pass the memories as the topic and these exact pipe-delimited columns:

Memory | Age | Why it exists

Order rows oldest-first — stale memories are the likeliest purge candidates and should lead.

Done when: the table is shown, one row per memory.

Step 4: Let the user choose what to purge

Present the memory names as a multi-select — the user ticks the ones to forget. Nothing is selected by default; an empty selection means "keep everything."

If the selection tool caps the number of options, batch the names across several prompts rather than silently dropping any. Always offer a "purge nothing" escape.

• Empty selection → report that nothing was purged and stop.

Done when: you have the confirmed set of memory names to delete.

Step 5: Purge and scrub

This is a hard delete — git history and the user's backups are the only safety net. For each selected memory:

1. Delete the memory file.
2. De-index — remove its line from MEMORY.md.
3. Unlink — find [[name]] backlinks to it in surviving memories and strip or neutralise them so no reference dangles.

Done when: every selected memory is gone and no trace of it remains in the store.

Step 6: Report

List what was purged (name + why) and what was scrubbed (index lines removed, backlinks cleaned). Keep it to a few lines — the deletions are the deliverable, not prose.

Done when: the user can see exactly what was forgotten.