Remember: Systematic Memory Management

Core principle: Memories are for retrieval, not storage. If you can't find it, you didn't remember it.

The Iron Law

code

NO MEMORY WITHOUT SEARCH FIRST. NO MEMORY WITHOUT VERIFICATION AFTER.

The Protocol

1. Search First

python

mem_search(query="<what you want to remember>", limit=5)

Found similar? Update or link to it. Don't create duplicates.

2. Write Specific Text

code

BAD:  "User likes Python"
GOOD: "Doctor Biz prefers uv for all Python package management. Never use pip, poetry, or easy_install."

BAD:  "Fixed a bug"
GOOD: "SQLite datetime gotcha: Use datetime(column) for comparisons, not raw column."

Include keywords you'd search for later.

3. Tag Consistently

Primary (pick one): project, preference, architecture, gotcha, solution, milestone, location

Secondary: project names, tech stack, status (project-complete, in-progress, deprecated)

4. Verify Retrieval

python

mem_search(query="<natural language query>", limit=3)

Can't find it? Fix the text or tags now. Don't move on until it's findable.

Quick Reference

Action	Command
Search	`mem_search(query="...", limit=10)`
Add text	`mem_add_text(text="...", tags=[...])`
Add file	`mem_add_file(path="...", tags=[...])`
Link	`mem_link(source_id="...", target_id="...", relation_type="...")`
Delete	`mem_delete(id="...")`

When to Remember

Yes	No
Project milestones	Temporary debug notes (use journal)
User preferences	Things findable via grep
Architecture decisions	Vague "might be useful" ideas
Debugging breakthroughs	Raw file dumps without context
API gotchas	Duplicates of existing memories

Red Flags

•"I'll search later" → Search now
•"Tags don't matter" → They're how you browse
•"Close enough" → Verify or it's lost
•"Might be useful someday" → Too vague, don't save it

Self-Correction

Created without searching? Search now. Delete duplicates.

Can't find it? Fix text/tags now. Don't move on.

Found conflicts? Resolve immediately. Use supersedes relation.

Session Workflow

code

Starting:  Search for project context, gotchas, architecture decisions
During:    Journal for reflection, note things worth remembering
Ending:    Create memories, verify all are findable