Memory Engineering — engineer the forgetting, not just the remembering

Memory Engineering is a skill for designing, reviewing, or pricing agent memory systems. It shifts focus from remembering to forgetting, providing scripts to price the write path, classify stored records, and enforce a forgetting policy. It produces cost analyses, architecture recommendations, audits, and policy linting results.

✨ What it does

  • Prices the write path with construction vs query split and cost per correct answer.
  • Scores four memory architecture families and names the cost each choice imposes.
  • Classifies records as FACT, SKILL, LOG, or PROSE and flags duplicates and staleness.
  • Lints forgetting policies with 8 checks, blocking on missing explicit rules and contradictions.
  • Refuses to return a 'best' system or auto-merge contradictions, surfacing findings instead.
  • Provides a worksheet and templates for designing a memory system with a forgetting policy.

🎯 When to use it

  • Adding memory to an agent and needing to choose between long-context, RAG, graph, or agentic memory.
  • Auditing what a CLAUDE.md or memory directory actually holds and deciding what to keep or expire.
  • When a memory store keeps growing with no clear policy on what leaves it.
  • When you need to justify the cost of a memory system with numbers, not just quality claims.

🚀 How to use

Trigger by asking to design, review, or price an agent memory system, or by providing a path to a memory directory or design spec JSON. The skill runs four stdlib scripts in sequence: price, pick, audit, and gate. Inputs can be a memory directory, a design spec JSON, or a question. Example prompts:

Price the memory system for my agent using this spec: workload.json
Audit my memory directory at ~/.claude/memory and tell me what to keep.

📄 Output: Cost reports, architecture recommendations, memory audits, and forgetting policy lint results, plus optional design documents and templates.

📦 Add this skill to Claude Code

# 1. Get the skills repo
git clone --depth 1 https://github.com/alirezarezvani/claude-skills /tmp/claude-skills

# 2. Copy this skill into your project (or ~/.claude/skills for all projects)
mkdir -p .claude/skills
cp -r /tmp/claude-skills/engineering/memory-engineering/skills/memory-engineering .claude/skills/memory-engineering

Skill source: engineering/memory-engineering/skills/memory-engineering/SKILL.md

⚠️ Good to know

Requires Python stdlib only, but the user must run scripts manually and interpret findings; no automated decision-making beyond exit codes.

❓ FAQ

What does the skill do if there is no forgetting policy?

The forgetting policy linter will fail with exit code 4, indicating a blocking issue, because a design without a forgetting rule is never considered done.

Can the skill auto-merge contradictory memories?

No, it never auto-merges contradictions; it surfaces them for the human to decide, as auto-merging destroys evidence of the conflict.

Does the skill provide a 'best' memory architecture?

No, it never returns a 'best' system; it names the cost the winning family makes you pay, and refuses to pick on a tie.

🤖 Overview, features, install steps and FAQ were generated from the project's SKILL.md on Sep 4, 2026. Always check the original source before running commands.