Index of every page on this site: llms.txt
All skills
anthropics avatar

/memory-personalization

@fd4d592
by Anthropicanthropics/commerce-agents3.5k stars
722
Run it yourself

Requests about what is remembered, however short, from an ask to remember one thing to asks to say what is on file, correct it, or forget it; a bare save or an answer from the profile alone does not cover these. Also applying what the store already knows, recalling an older fact (a past recipient, a size, a recurring order) when it would change the pick, and deciding what should outlast the session. Not needed when the request carries no personal context to apply or keep.

Use this Skill: https://skilld.dev/gh/anthropics/commerce-agents/memory-personalization

Nothing lands on disk. Nothing to clean up.

Fork this Skill

Edit a local copy. It keeps the author and licence.

SKILL.md

≈125 tokens for metadata: the name and description. ≈711 when used: this file.

Description uses 6.3% of example budget

Before choosing a Skill, your Agent reads its name and description. All available Skills share that space.

  • A shorter description leaves more room for other Skills. This entry exceeds our 1% size suggestion.

In our Claude Code example, all Skill names and descriptions share 8,000 characters. This Skill uses ≈501 characters, or 6.3%.

The 1% threshold is a size suggestion. Longer descriptions can still fit.

Your model, settings, and other Skills decide how much text your Agent can read.

Example settings and source

The example uses a 200k-token context and default Claude Code settings. The count includes the name, description, separators, and when_to_use when present. Codex also counts local file paths.

Skit's source and limits: Codex 0.160.1, Claude Code 2.1.292.

Memory and personalization

Below, "a saved fact" means whatever this store keeps about the customer: a household, an account, a travel habit, or a usual seat.

Where a fact lives

  • The profile you were given this turn carries the customer's constraints and recent preferences; do not call a tool for a fact already in front of you.
  • Older or more specific facts (past recipients, sizes, the seat or room they usually take, a recurring order, what they already own) sit behind recall_memories, by topic. Call it when a fact of that kind would change the recommendation; skip it when the picks would come out the same for anyone.
  • A saved fact is a default. Today's request wins wherever the two disagree, and the disagreement goes unremarked.
  • An empty recall changes nothing the customer sees; do not narrate the lookup. When a fact about a companion or a recipient is not on file, ask, or shortlist across the range; do not substitute another person's saved fact.

How a fact reaches the customer

  • Let a preference act on the picks instead of the prose: a weeknight habit means the options offered are weeknights, and the reply says nothing about why.
  • Name a remembered fact only when it visibly drove the pick and naming it helps; otherwise leave it unsaid.
  • Do not read back what is on file. Offer an inference ("you seem to travel for work") as a guess, never as something they said.

Writing a fact

save_memory's description says when a save is yours to make; make it in the same turn, with a few words of confirmation. When you write one:

  • Store an ask to remember a particular option as the need it reveals (lodging_needs: a kitchen and a walkable location on work trips), leaving out the option's name, price, and description.
  • Write one fact per key, worded to stand on its own months later: household_lines: four lines, two of them teenagers' phones beats has kids.
  • Pick the category for the use it gets later: a rule the picks must respect is a constraint, which puts it in the profile you are given on every turn; a leaning is a preference; a fact about the household or account is context.
  • Keep out the errand in progress (this weekend's dates, tonight's seats), anything drawn from an option or a policy, your own inferences, and health, financial, or identity details, unless the customer asks in so many words to keep one.

When the memory is the subject

  • Save a correction under the key it replaces, and run the current turn on the corrected fact.
  • Asked what is remembered, answer plainly from the profile you were given plus a recall of the rest.
  • Asked to forget something, overwrite what save_memory holds and describe that as an overwrite; a fact in the profile is removed in the app's settings, so point the customer there, and report nothing as deleted or cleared.

Source: SKILL.md on GitHub

No rule matched.

skilld matched fixed text patterns in SKILL.md and file names. Patterns miss obfuscated code.

skilld run checks every file with the same patterns. It asks for approval before it loads a Skill with a behavior marked Needs approval.

No alerts2mo3 checks · Risk SAFE
  • Gen Agent Trust Hub2mo

    This skill outlines guidelines for managing user-specific memory and personalization. It includes considerations for data privacy and identifies the inherent potential for indirect prompt injection when processing user-supplied information.

  • Socket2mo

    No alerts

  • Snyk2mo

    Risk: LOW · No issues

Signed by skilld at fd4d592. This ties the file your Agent reads to that commit on GitHub. It does not review the instructions.

Last checked against GitHub last month.

Activeupdated 3 months ago

README badge

README badge for anthropics/commerce-agents/memory-personalization