What does Claude Code's summary keep when a session runs long?

Why this matters to you: when a long Claude Code session fills up and gets summarized, the agent keeps going on the summary. If it had loaded a skill, this is how much of that skill's wording we found in the summary.

The short answer: very little of the skill's own wording. In 42 summaries from 15 of our sessions, we found 28 times an agent had loaded a skill before a summary, and searched each summary for distinctive phrases from that skill. Across those 28 skill loads, about one distinctive phrase in 700 turned up word for word (39 of 28,085). In 21 of the 28, no phrase from the skill appeared at all. Reading the same summaries, we did find the messages people had typed and the files being worked on.

What we found

• Phrases: across 28 skill loads in 15 sessions, 39 of 28,085 distinctive five-word phrases appeared in the following summary word for word (0.14%, about one in 700).
• Skill loads: 21 of the 28 skill loads (75%, three in four) had no matching phrase in the summary.
• Skill names: the skill's name appeared in the summary in 5 of 28 cases (18%, about one in five), including when the skill itself was the subject of the work.
• The summaries: all 42 used the same nine-section layout and ran about 5,250 tokens on our estimate (from about 3,640 to 7,120), several pages of text.
• What we saw them keep: reading the summaries, we found the messages people had typed (often word for word, typos included), about nine file paths per summary (8.7 on average across the 42), errors with their fixes, and the tasks still to do. The messages, errors and tasks are observations from reading; the average of 8.7 file paths across the 42 summaries is a count.
• Looser matching: across the 28 skill loads, counting shared key words instead of exact phrases found 10% to 27% overlap, roughly one to three words in ten. Our skills share a lot of vocabulary with our conversations, which inflates this number, so it doesn't tell us how much meaning carried over.

How we tested it

• When: August 1, 2026.
• Sample: every one of our Claude Code session records that contained a summary at the time: 15 sessions with 42 summaries, from Claude Code 2.1.209 through 2.1.220.
• Skill loads in scope: 28 cases where an agent loaded a skill's full text before a summary, covering 8 different skills and about 55,800 estimated tokens of skill text in total, about 2,000 per load.
• Scoring: each skill's text was split into five-word phrases. A phrase counted as distinctive when at least one of its words had six or more letters and was not a common word. We then searched the summary for each distinctive phrase, word for word.
• Tokens: estimated as characters divided by four, not counted with a tokenizer.

Limits

• Small sample, one team. 15 sessions, 28 skill loads and 8 skills, all ours, and our skills share a lot of vocabulary.
• Word for word misses paraphrase. A summary can carry a rule in different words, and this count can't see that. The 75% figure has the same limit: it shows that in 21 of 28 skill loads the summary had no exact phrase from the skill, not whether they paraphrased it.
• Claude Code adds some skills back. We found 21 such attachments, in 9 of the 15 sessions, where Claude Code attached the full text of loaded skills to the new context right after a summary, outside it. It did not do this for every skill: in at least two sessions, a skill the agent had just used was left out. This study measures the summary only.
• Versions. Later Claude Code versions may summarize differently, and we did not study Codex.

What we changed

We stopped relying on the summary for anything that has to survive. We ask the agent to write a short checkpoint before a summary: the goal, the next action, the skills to reload and the rules already settled. In the engram agent kit v0.2.0, a hook asks once for a fresh checkpoint before an automatic summary, and when a checkpoint exists, another hook puts it back into the new context word for word, up to 6,000 characters. Step 6 of the lesson sets this up and shows how to check it worked.