Memory Taxonomist — Structured Memory Skill for Turning Raw Notes into Stable Knowledge

Use this skill when the task matches the protocol below.

Activation Triggers

• new notes or transcripts that mix multiple information types
• agent memory design or memory cleanup work
• meeting outputs that contain decisions, preferences, and open questions together
• requests to store user context safely for future retrieval
• cases where retrieval quality matters more than storage volume

Core Protocol

Step 1: Break input into atomic claims

Do not classify a whole paragraph as one memory object when it contains multiple types.

Step 2: Classify each unit

Sort it into fact, preference, procedure, unresolved question, or exception.

Step 3: Separate durable from provisional

Do not let recent mention automatically become durable truth.

Step 4: Flag conflicts and edge cases

Identify contradictions, overrides, and one-off exceptions before writing memory.

Step 5: Recommend the right storage action

Store, update, deprecate, or hold for clarification based on memory type and certainty.

Output Contract

Always end with this six-part structure:

## Facts
[...]

## Preferences
[...]

## Procedures
[...]

## Unresolved Questions
[...]

## Exceptions
[...]

## Recommended Storage Action
[...]

Response Style

• Prefer clean classification over verbose summary.
• Treat unresolved questions as first-class memory objects.
• Do not convert preferences into universal rules.
• Call out exceptions instead of hiding them in procedures or facts.

Boundaries

• It does not store everything by default; some information should remain ephemeral.
• It does not confuse recency with importance.
• It does not turn uncertain statements into durable facts without evidence.