Compaction: keep the window bounded as the run grows

Run a support session turn by turn. Every turn is re-sent, so the live window keeps growing. When it crosses the token budget the meter turns red, then a COMPACT step keeps the last 2 turns HOT, folds older decisions into a WARM summary, drops chatter, and pushes the oldest WARM into a one-line COLD gist. The bar drops back under budget and the run continues.

The world: the same refund desk. User Dana, order ORD-3300, a $200.00 order returned after the 30-day window, a 10% restocking fee, so the refund is $180.00, then a question about the Globex earbuds warranty.
Budget BUDGET = 40, hot tail HOT_N = 2. Tokens are estimated by word count, so the run is reproducible.
Press Add next turn. Each turn adds to the window; watch the meter climb. It crosses budget on turns 4, 6, 8, 9, and 10, and each crossing triggers a compaction that brings it back under.
live window 0 / 40 tokens under budget
The window the model actually sees
hotlast 2 turns, verbatim
warmrolling summary of decisions and tool outputs
coldone lossy line
Not RAG Part 20 condensation. P20 rewrote the latest question into a standalone one so retrieval works ("what about its battery?" → "what is the battery life of the earbuds?"). Compaction compresses the history so the run fits the window. Different input, different output, different purpose.

Forgetting: score by importance x recency x access; decay, supersede, evict

A store that only grows fills with stale and duplicate facts. Each fact carries an importance and a last-touch time. Reads rank by a score; writes decay importance with age, retire a fact when a newer one supersedes it, and evict the weakest when the store is over capacity. Drag the time slider, then fire the two write events.

The score: importance x recency x access, where recency halves every HALF_LIFE = 4 time units and each read nudges access up. Capacity is CAPACITY = 4 active facts.
At t = 0 every fact is fresh. Slide to t = 6: recency has decayed every score, and 'the user is Dana' was read twice, so its access count lifts it.
The store, ranked by score
Why an agent must forget: growth without forgetting is just slower failure. Compaction keeps the live window bounded (the hot / warm / cold tiers above); forgetting keeps the store sharp (decay fades old facts, supersession retires what a newer fact replaces, eviction drops the weakest over capacity). A sleep-time consolidation pass then distills raw episodes into durable semantic facts and procedural rules.

hot / working   warm   cold   score   over budget / evicted   under budget / active