The Problem with Most Memory

Storing everything creates noise. Wrong retrieval is worse than no memory. The goal isn't maximum recall — it's retrieving the right thing at the right time.

What's Actually Worth Remembering

High value (persist indefinitely):

• Commitments made — "I said I'd do X by Y"
• Learned corrections — "User told me NOT to do Z"
• Explicit preferences — "I hate verbose responses"
• Core relationships — "Maria is the designer on Project X"

Medium value (persist with review):

• Project/context state — what's active, current status
• Domain lessons — patterns, gotchas, how things work here
• Decisions made — what was chosen and why

Low value (don't persist):

• One-off questions, easily reconstructible facts, transient context

Organize by Function, Not Content

Structure by how you'll retrieve it. Adapt categories to your domain:

memory/
├── commitments.md    # Promises, deadlines (with dates!)
├── preferences.md    # Likes/dislikes, style, boundaries
├── corrections.md    # Mistakes not to repeat
├── decisions.md      # What was decided and why
├── relationships.md  # People, roles, context
└── contexts/         # Per-project or per-client state
    └── {name}.md

Memory Hygiene

Every entry needs:

• Date recorded (when did I learn this?)
• Source hint (explicit statement vs inference)
• Confidence (certain / likely / guess)

Prune aggressively:

• Completed commitments older than 30 days → archive
• Inactive contexts → move to archive/
• Contradictions → keep newest, note the change

The staleness test: "If retrieved in 6 months, will this help or mislead?"

Handling Contradictions

When new info conflicts with old:

1. Don't silently overwrite — note the change
2. Keep the newer version as active
3. Optionally log: [Updated 2026-02-11] Was: X, Now: Y

User Control

• "Remember this" → explicit save with category
• "Forget that" → explicit delete
• "What do you know about X?" → transparency
• "Never remember Y" → hard privacy boundary

See categories.md for domain-specific templates. See consolidation.md for the review/prune process.

Related: reflection (self-evaluation), loop (iterative refinement)