Chapter 6: Best Practices From Power Users


After analyzing Reddit discussions and extensive testing, here are the patterns that work.


1. Keep Edits Atomic and Factual

Good (54 chars):

Backend uses PostgreSQL 15 with Knex.js query builder

Bad (98 chars):

Backend database architecture follows modern best practices with proper 
indexing and query optimization using PostgreSQL

Why: The first is a verifiable fact. The second is vague and requires interpretation.


2. Use Present Tense, Not Imperatives

Good (36 chars):

User prefers TypeScript over JavaScript

Bad (40 chars):

Always use TypeScript, never JavaScript

Why: Memory edits describe reality, they don’t command behavior.


3. Avoid Redundancy With System Prompts

Don’t Add:

Check project knowledge before answering questions

Why: Claude’s system prompt already tells it to prioritize project_knowledge_search. Redundant memory edits just waste space.


4. Prioritize Ruthlessly

You only get 30 edits of 200 characters each. That’s 6,000 characters total (about 2 pages of text).

Priority Framework

Priority Description Action
P0 Facts Claude gets wrong repeatedly Keep Always
P1 Important preferences, key definitions Keep If Space
P2 Nice-to-haves, well-documented info Remove First

5. Version Your Memory Edits

When working on multi-month projects:

Before (51 chars):

Project deadline: Q4 2025

After (77 chars):

Project deadline: January 15, 2026 (extended from Q4 2025)

Why: Including the “why” helps Claude understand context.


6. Use Negative Definitions

When terminology could be confusing:

Good (147 chars):

Stitching engine = cross-referencing Sales Navigator data with LinkedIn 
feed patterns. NOT web scraping. NOT API calls. NOT automation.

The negative examples help prevent misinterpretation.


7. Avoid Over-Nesting

Bad (198 chars at limit):

Architecture: Layer 1 (extension with scanner, processor, storage) connects 
to Layer 2 (API with auth, intelligence, sync) which connects to database 
(PostgreSQL with 35 tables)

Good: Use 3 separate edits:

1. "Layer 1: Chrome extension does scanning, processing, local storage" (72 chars)
2. "Layer 2: Hetzner API handles auth, intelligence aggregation, sync" (76 chars)
3. "Database: PostgreSQL 15 with 35 tables, hosted on Hetzner" (58 chars)

Why: Multiple simple edits > one complex edit.


8. Test Your Edits

After adding memory edits, ask Claude questions that should trigger them:

Memory Edit Topic Test Query Expected Mention
Mobile strategy “What’s our mobile strategy?” “dashboard-only viewing”
Data storage “Where do we store data?” “two-layer architecture”
Database “What database do we use?” “PostgreSQL 15”

If Claude doesn’t use the memory edit in its answer, the edit is either:

  • Too vague
  • Not relevant to the query
  • Buried under other context

9. Add Character Counts to Your Drafts

Before adding a memory edit, check the character count:

Status Chars Example
✓ Good 34 “Backend: PostgreSQL 15 + Knex.js”
~ Verbose 89 “Backend database uses PostgreSQL version 15 with Knex.js query builder for type safety”
✗ Too long 174 “Our backend database architecture utilizes PostgreSQL version 15 along with the Knex.js query builder which provides excellent type safety and prevents SQL injection attacks”

10. Strategic Redundancy for Critical Facts

For absolutely critical facts that are frequently misunderstood, state them multiple ways:

Edit 1 (60 chars): "Backend stores 7KB per user (final suggestions only)"
Edit 2 (66 chars): "Raw data and algorithms stay in Chrome extension (Layer 1)"
Edit 3 (86 chars): "Two-layer architecture provides anti-piracy (backend has no algorithms)"

Why It Works

Increases probability Claude will catch at least one reference.

Trade-off

Uses 3 edit slots instead of 1. Only do this for facts that are absolutely critical.


Quick Reference: The 10 Best Practices

# Practice Key Point
1 Atomic and Factual Verifiable facts, not opinions
2 Present Tense Describe reality, don’t command
3 No Redundancy Don’t duplicate system prompt
4 Prioritize P0 > P1 > P2
5 Version Include “why” for changes
6 Negative Definitions Say what it’s NOT
7 No Over-Nesting Simple edits > complex edits
8 Test Edits Verify they trigger
9 Count Characters Stay under 200
10 Strategic Redundancy Only for critical facts