An agent that reads and updates a bounded, domain-owned mental model across sessions becomes measurably better at session 50 — a stateless generalist is the same agent every time.

Principle

Specialization Compounds

The wager

Session 50 is the tell.

Two agents. Same codebase. Fifty sessions.

domain skill →s1s25s50

— Stateless generalist Rediscovers conventions every run. Session 50 agent ≡ session 1 agent. Every run pays the cold-start tax.

— Specialized + updatable Reads a bounded model, appends what it learned. Carries 50 sessions of domain learnings. This is the compounding.

Without state, the agent is tired at session 1 and tired at session 50.

The mechanism

A mental model works only when all three hold.

① Bounded

~200-line budget. Oldest summarizes.

Loads in hundreds of tokens, replaces thousands of rediscovery.

Miss this → 2,000 lines of lore, noise drowns signal.

② Updatable

Agent appends after non-obvious learnings.

Not every task updates. Only the ones where something was actually learned.

Miss this → read-only snapshot goes stale fast.

③ Domain-owned

One role, one file, one owner.

backend-mental-model.yaml doesn't carry frontend context. Role is the boundary.

Miss this → specialization inverts into generality.

Drop any one property and the file stops being memory — it becomes noise, staleness, or cross-contamination.

The artifact

One file, five sections, one owner.

expertise/backend-mental-model.yaml

# role: backend-dev · owner: /backend-dev agent · budget: 200 linesconventions:stable-"migrations always go through just migrate"-"API responses use snake_case, never camelCase"gotchas:hard-won-"auth middleware runs AFTER CORS — order matters"-"sessions cache bypasses Redis on staging"architectural_decisions:load-bearing-"we chose pgvector over Pinecone — don't revisit"preferred_patterns:opinionated-"use Result[T, E], not raise — errors are data"known_debt:ack'd-"jobs/worker.py is a mess — scheduled for rewrite"

Five sections loads in hundreds of tokens — and replaces thousands of tokens of rediscovery the agent would otherwise pay every session.

What breaks it

Four failure modes. Each kills the compounding.

✗

Stateless (no file)

Every session is session 1. Agent relearns conventions, reinvents the auth pattern, guesses at gotchas. Quality ceiling is bounded by cold-start.

✗

Read-only (never updated)

Memory without learning. File is a snapshot of an earlier state. Agent reads it, notices it's wrong, doesn't write back. Staleness compounds instead of skill.

✗

Unbounded (no line budget)

Noise dominates signal. By session 100, 2,000 lines of accumulated lore. Agent's context budget is spent reading the memory, not doing the work.

✗

Shared across roles

Specialization inverts. Backend's file accumulates frontend details. Each role carries both domains. Agents get more generic with each update — not more specialized.

The file is memory only when all four traps are avoided — miss one and compounding stops.

The discipline

One role, one file, one owner, one line budget.

Institutional memory is a shape, not a hope. Agents that write back become better engineers in their domain with every session. Agents that don't stay tired forever.