Memory, Skills, Rules: The Compression Spectrum for Agent Operators

Agents die in two ways: they forget, or they drown in context.

The fix is not "bigger window forever." It is compression: turn experience into the right artifact at the right density.

This guide gives operators a simple spectrum: memory → skills → rules.

Kill the default approach

Default: paste the whole chat history into every session, or invent twenty agents that each re-discover the same procedure.

What breaks first: token cost, contradictory notes, and "we already decided this" loops.

By the end you will have

1. The compression spectrum (what goes where)
2. The Experience Shelf Layout magnet
3. Promotion rules (when memory becomes a skill)
4. Failure modes
5. A weekly prune ritual

The spectrum

Think of three densities:

Memory is cheap to write and expensive to re-read.
Skills are medium.
Rules are few and sharp.

Promotion ladder

1. You did a thing once → journal memory (receipt)
2. You did it three times the same way → skill with verify
3. Breaking it always hurts → rule in AGENTS.md / policy

Do not jump from one incident to a permanent rule unless the blast radius is high.

Magnet: Experience Shelf Layout

Name: Experience Shelf Layout

repo/
  AGENTS.md           # hard rules (short)
  skills/             # procedures with verify
  journal/            # dated receipts (memory)
  scripts/check.sh    # green means green

Journal template:

# 2026-07-14-mint-sim
 
## Decision
Simulate before any mint-related tx.
 
## Evidence
Dry-run caught bad ATA assumption.
 
## Next
Encode sim step into skill: solana-tx-sim

You should see: a journal entry promoted to a skill after the third repeat, with a verify command that fails when skipped.

How this maps to multi-agent hype

Multi-agent systems often re-discuss what should already be a skill or rule. That is inverse compression: more tokens, less memory.

Prefer:

• One agent
• Skills for how-to
• Rules for never
• Journals for what happened

Use parallel agents only for parallel research, then write a single memory receipt.

Weekly prune (15 minutes)

[ ] Delete journal noise older than 30 days (keep decisions that still bind)
[ ] Skills unused in 30 days: archive or delete
[ ] Rules: still true? still needed?
[ ] check.sh still green?

Failure modes

When not to over-compress

• You are exploring a brand-new domain (stay in journal mode)
• Legal/compliance needs full audit trails (keep raw logs separately)
• The procedure still changes every run

Path options

Bottom line

Compress on purpose. Do not hoard chat.

Your next action: create Experience Shelf Layout in one repo and promote one repeated procedure into a skill with a verify step.