Context Compaction and Conversation History Management

Most chat APIs are stateless. Every request resends the conversation so far, and the model reads it from scratch. Tokens, cost and latency therefore grow as the session gets longer, until the history hits the context limit and the request fails or gets truncated. Quality can drop even before that point, because relevant details are buried among old material. The simplest strategy is truncation, usually as a sliding window: keep the system prompt and the most recent turns, and drop the rest. It is cheap and predictable, but it forgets early facts, such as constraints the user stated at the start. A common refinement is to pin important messages so they are never dropped. Rolling summarization replaces older turns with a model-written summary that is updated as the conversation grows. It keeps the gist in far fewer tokens. The cost is that summaries lose detail and can introduce errors, and those errors compound when a summary is itself summarized again. Compaction is a more deliberate version used in agent tools. When token usage crosses a threshold, the system asks the model to write a structured handoff covering the goal, progress, decisions, open problems and key file names. It then starts a fresh context with that handoff plus the most recent turns. Claude Code, for example, has a compact command and can also compact automatically near the limit. Tool-result clearing targets the biggest source of bloat in agent sessions: large outputs from searches, file reads and commands that were useful once. Swapping them for short placeholders frees space and still records that the call happened. External memory stores facts in a file or database and retrieves them when relevant, so they survive any trimming. A common misconception is that the model remembers earlier parts of a chat on its own. Any continuity comes from what the application chooses to resend or retrieve.