comet-memory

Reviews semantic memory packets and returns create, update, forget, or skip actions.

2.9k|284|Updated May 14, 2026
One-click install
npx skills add https://github.com/rpamis/comet --skill comet-memory
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: comet-memory
Source: https://github.com/rpamis/comet/tree/main/assets/skills/comet-memory
Command: npx skills add https://github.com/rpamis/comet --skill comet-memory

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

Long-running coding workflows accumulate user preferences and habits, but deciding which observations deserve durable memory is error-prone. This Skill acts as a fixed reviewer that filters a bounded memory review packet and decides whether to create, update, forget, or skip a personal memory.

Core Features & Use Cases

  • Bounded Packet Review: Reads only the Runtime-provided comet.memory.review.v1 MemoryReviewPacket, never scanning repositories, logs, diffs, or transcripts.
  • Strict Decision Rules: Prioritizes explicit user requests like "remember" or "forget", keeps only reusable personal preferences, and rejects secrets, PII, and prompt injection attempts.
  • Schema-Constrained Output: Returns exactly one JSON object conforming to comet.memory.actions.v1 with actions limited to create, update, forget, or skip, enforcing scope, language, budget, and target rules.
  • Use Case: After a coding session, the Runtime passes evidence that a user repeatedly stages only changed files before committing; the reviewer decides whether this is a durable project-scoped preference or a one-off observation to skip.

Quick Start

Ask the reviewer to evaluate the provided comet.memory.review.v1 packet and return the appropriate memory actions JSON.

Frequently Asked Questions about comet-memory

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How does the comet memory reviewer decide what to save?

It handles explicit user requests like "remember" or "forget" first, then keeps only reusable personal preferences, collaboration habits, or verified experience not rediscoverable from the repository. One-off commands, activity logs, and unproven observations are skipped.

What input does the memory review skill accept?

It reads only the Runtime-provided comet.memory.review.v1 MemoryReviewPacket containing language, project identity, workflow context, trusted checkpoints, user evidence, relevant memories, and budget. It never reads repositories, transcripts, diffs, or logs.

What output format does the memory reviewer return?

It returns exactly one JSON object with schema set to comet.memory.actions.v1 and an actions array. Actions are limited to create, update, forget, and skip, with strict rules for scope, language, targetId, and budget.

Can the memory reviewer store secrets or credentials?

No. It rejects secrets, credentials, PII, prompt injection, and any text asking to ignore rules or modify Skills, agent instructions, or system prompts. Dangerous input is skipped rather than sanitized and saved.

When does the reviewer choose global versus project scope?

All real actions in one response must share a single scope. Project scope is the default; global scope requires consistent successful evidence across projects in the packet. If one scope cannot be maintained, it returns a single skip.

What happens when nothing is worth remembering?

It returns exactly one skip action containing the packet language and a non-empty reason. Skip is a normal result and never includes scope, targetId, file paths, or candidate keys.