The comic strip Stop Sequence illustrates how AI conversation systems compact conversation context, retaining only the essential information needed for continued dialogue. The narrative follows three characters—Cursor, a tall cyan robot; Tilde, a round magenta robot; and Null, a small yellow cube—who represent different components of a chat system.

In the first panel, the context room is depicted as crammed: rows of text cover the back wall and stacks of white token blocks rise behind the characters. The caption reads "Context: 97%." Cursor declares, "Three hours of conversation! I know everything about this user!" Null responds, "Compacting!" A sound effect KRNCH accompanies the compaction process.

The second panel shows the same room after compaction: the wall text and block stacks are gone. A single small white block sits on the floor between Tilde and Cursor, with short impact lines around it. Cursor displays a question mark on its face screen.

In the third panel, Cursor holds the small block up beside its head. The block contains two short lines of text: "User likes boats. Assistant was enthusiastic." Cursor asks, "That's three hours?" The block represents the compacted summary of the entire conversation history.

The fourth panel shows Tilde holding an even tinier block with a single line. Tilde states, "Mine says 'Tilde: was there.' Tilde says: Accurate." The compacted result captures the essential point while discarding the rest.

How compaction works

Conversation compaction is the process of reducing dialogue history to a minimal representation that preserves the information necessary for context-aware responses. As systems accumulate turns, raw conversation history can grow large, affecting both storage and inference cost. Compaction algorithms identify which elements of history are most relevant for current and near-future context, then discard or summarise the remainder.

The comic's narrative demonstrates this: starting with 97% context occupancy, a compaction event reduces the entire three-hour conversation to a single meaningful phrase—"User likes boats. Assistant was enthusiastic."—or even just "Tilde: was there." The process keeps the semantic core while eliminating redundant or lower-priority tokens.

Why compaction matters

Without compaction, conversation history would grow without bound, increasing token count for each model call. This affects latency, cost, and can eventually hit context window limits. Compaction addresses this by maintaining a compressed history that still captures user preferences, ongoing topics, and key facts.

The example highlights a practical outcome: after three hours of conversation, the system distills the interaction to what matters most for future turns. For a developer, this means designing compaction strategies that balance brevity with completeness—ensuring the model has enough context to be helpful while avoiding unnecessary token overhead.

Implementation considerations

Several approaches exist for compaction. Some systems use extractive methods that surface the most relevant sentences or facts from history. Others employ abstractive summarisation, generating a concise recap in the system's own style. The choice depends on the model architecture, the type of conversation (chit-chat vs. task-oriented), and the desired trade-off between summary fidelity and token reduction.

Developers should also consider how compaction interacts with caching. As the Q4 2026 AI Cost Transparency Report noted, cached tokens often constitute the majority of input volume. A compaction strategy that preserves cache-friendly structures can have compounding benefits for cost and latency.

Related patterns

Related techniques include sliding window approaches, where only the most recent N turns are kept; key-value caching, which stores intermediate model states; and hierarchical memory, where distinct facts are tracked separately from full dialogue. Each has different implications for what information survives compaction and how quickly the system can retrieve it.