Memory Retention (prune)

claude-mem-go's prune subcommand deletes observations older than a cutoff. This is the ONLY destructive operation this project's CLI exposes — treat it with real caution, not like mem-search/mem-doctor.

When to use

Only when the user explicitly asks to delete or clean up old memory — never proactively, and never as a side effect of some other request. "The database seems large" is not, on its own, a request to delete anything; ask first.

The safety rule: dry-run first, always

prune defaults to a dry run — without -yes, it only reports how many observations WOULD be deleted, and deletes nothing. Always run the dry run first, show the user the count, and get their explicit confirmation before ever adding -yes. Do not skip straight to -yes even if the user's request sounds specific ("delete everything older than 90 days") — they may not have known how many observations that actually covers.

# Step 1 — always this first, and show the result to the user:
"$CLAUDE_PLUGIN_ROOT/claude-mem-go" prune -older-than-days 90

# Step 2 — only after the user confirms the count is what they want:
"$CLAUDE_PLUGIN_ROOT/claude-mem-go" prune -older-than-days 90 -yes

# report, then strip, relative file paths file-context can never match
"$CLAUDE_PLUGIN_ROOT/claude-mem-go" prune -relative-paths
"$CLAUDE_PLUGIN_ROOT/claude-mem-go" prune -relative-paths -yes

(Fall back to claude-mem-go prune ... on PATH if $CLAUDE_PLUGIN_ROOT isn't set — see mem-doctor's skill for why.)

Confirm WHICH store you're about to delete from

prune acts on one store, and which one is not obvious. With no -db flag it uses $CLAUDE_MEM_DB when that is set (the same store the hooks and MCP server write to), and the local SQLite file otherwise.

That matters more here than anywhere else, in both directions. A dry run against the wrong store reports a count that has nothing to do with the data the user is thinking of — usually 0, which reads as reassuring right up until -yes deletes from somewhere else entirely. Run mem-doctor first and show the user the store it names, so the count they approve is a count from the database they mean:

"$CLAUDE_PLUGIN_ROOT/claude-mem-go" doctor   # names the store, and says
                                             # whether $CLAUDE_MEM_DB chose it

Pass -db <postgres DSN> explicitly if it needs to be a different store than the one doctor reports — and if you do, say so to the user, because it is then not the store their hooks are writing to.

Flags

• -older-than-days N — required; there is no default (an unset cutoff refuses to run rather than guessing).
• -relative-paths — a different mode, mutually exclusive with -older-than-days: strips RELATIVE entries from files_read/ files_modified on observations written before file paths were canonicalized. The PreToolUse file-context lookup can never match those (Claude Code sends absolute paths), and they cannot be made absolute — the cwd they were relative to was never stored, and a guess would make the row match the wrong file. The observations are kept; only the unmatchable path entries go. Dry-run without -yes.
• -db <path-or-DSN> — the store to prune. Defaults to $CLAUDE_MEM_DB if set, else the local SQLite file. See the section above.
• -project name — scope to one project instead of every project in the store. Ask the user which they mean if it's ambiguous ("my memories" vs "all projects").
• -yes — actually delete. Omit it for the dry run.

Before suggesting this at all

If the user's real concern is disk usage or a very large database, consider whether export (see the mem-export skill) — backing up to a file — is what they actually want instead of, or before, deletion.