Memory Distiller

---

name: memory-distiller

description: Distill repeated user preferences, successful patterns, and durable working rules into reusable memory notes or prompt-ready context blocks. Use when a user wants to capture habits, preserve preferences, summarize lessons from prior work, or convert raw conversation/task outcomes into structured memory.

---

# Memory Distiller

Overview

Use this skill when the user wants to turn raw interaction history into stable,

reusable memory. The goal is not to summarize everything. The goal is to keep

only the parts that are durable enough to improve future work.

Read `references/output-format.md` when the user wants a structured output

template, a prompt-ready context block, or a reusable memory profile format.

Read `references/example-prompts.md` when the user needs prompt examples,

variation ideas, or help choosing the right invocation pattern.

Quick Start

If the user does not specify a format, default to this flow:

1. extract candidate memories from the source material

2. keep only durable and evidence-backed items

3. rewrite them as future-facing rules

4. return:

- stable preferences

- working rules

- anti-patterns

- one short reusable context block

If the user already has a memory document, switch into review mode instead of

rebuilding everything from scratch.

When To Use

Use this skill when the user asks to:

Do not use this skill for:

Output Selection

Choose the narrowest output that matches the user's goal:

- use when the user wants a compact long-term preference document

- use when the user already has notes and wants to remove weak items

- use when the user wants a short block to reuse in future prompts

- use when the user wants to know what should be kept, rewritten, or removed

Read `references/output-format.md` before producing any structured output.

Core Rule

Only preserve information that looks durable.

Good candidates:

Weak candidates:

When a memory candidate is uncertain, mark it as tentative or exclude it.

Evidence Threshold

Prefer memories that are supported by one of these:

Prefer to exclude items that are supported only by:

Workflow

1. Gather source material

Start from the material the user provides or points to:

If the source material is large, first compress it into candidate signals rather

than copying everything forward.

2. Extract candidate memories

Look for statements that imply stable behavior, such as:

Group candidates into a small set of categories:

When possible, tag each candidate mentally as one of:

3. Remove weak or noisy items

Drop any item that is:

Prefer precision over recall. A small memory set with strong signal is better

than a large noisy list.

4. Rewrite into future-facing rules

Rewrite valid items as clear, reusable guidance.

Prefer forms like:

Avoid forms like:

5. Produce the requested output

Choose the narrowest useful output for the user:

If the user does not specify a format, default to:

1. Stable preferences

2. Working rules

3. Anti-patterns

4. A short reusable context block

Examples

Example: conversation to profile

If the source says:

The distilled result should look like:

Example: task outcomes to rules

If repeated successful tasks show:

The distilled result should look like:

Example: weak candidate to exclude

If the only evidence is:

Do not convert that into a durable preference unless there is more support.

Output Guidance

When producing memory content:

If a prompt-ready context block is requested, keep it short enough that it can

realistically be reused without bloating future prompts.

Safety And Quality