A smaller model rewrites the middle of a Claude Code session at about 160K tokens

Plan around the retained tail. A rule typed at turn 12 of a 90-turn session is outside it, so the rule survives only if the summarizer chose to write it down.

Roughly 40,000 tokens sit between the compaction trigger and the window ceiling ^[15]. Compaction fires when the count crosses, and the agent continues from the reconstructed buffer of system prompt, summary and recent turns without signalling that anything happened ^[7].

The summary aims at outcomes: files modified, errors resolved, dependencies added ^[5]. The reasoning that produced them is not in scope, so a session where the agent tried four approaches before one worked compacts to "implemented authentication using JWT" with no record of the three rejected strategies or why they failed ^[6]. Nothing left in the buffer contradicts the agent if it proposes one of those three again. The post describes the resulting failure as syntactically correct code that violates a constraint the agent learned 50 messages earlier and forgot during compaction ^[9].

Branching sessions lose the most. Parallel exploration collapses into a single narrative, and the distinctions between the approaches go with it, while linear conversations where each exchange builds on the last summarize coherently ^[11].

The post recommends putting constraints, decisions and open tasks in persistent artifacts, because sessions that create explicit artifacts produce summaries that reference those artifacts ^[10]. A file path costs a handful of tokens in the summary. The bytes it points at were never in the conversation buffer, so no summarization pass can thin them.

These figures come from one runtime. For 160K to be your trigger you need the same 200K window and the same 80 percent threshold in the build you are running ^[1]. The three-phase description is not sourced to Anthropic documentation in the post, which states it was written with the assistance of AI under human supervision and review ^[14]. The retention rule transfers. If your runtime keeps the system prompt and the recent turns and rewrites everything between them, anything you need at turn 90 belongs on disk.

The post's other two levers are manual compaction, which it says gives control over what survives by forcing early summaries with explicit retention rules ^[12], and token discipline: concise system prompts, artifact-based state and pruning dead branches, which it says delays or eliminates compaction in most sessions ^[13].