61 days of Claude Code transcripts: hooks reported success but only 11% of context reached the model

JhouHate · reddit · 2026-09-30

The author built Hipocampo, a system around Claude Code's native features (hooks, MEMORY.md, CLAUDE.md, skills, subagents, compaction — no installs, no MCP) to control what enters each session's context window, and measured it over 61 days of transcripts (190 main sessions, 918 subagent files).

Key findings:

Organizing model: Resident (always delivered before first decision), Paged (only when opened), Interrupt (only on matching events).
0 injected recalls in 190 conversations — memory only entered when the agent opened its file; 209 of 432 memories were opened at least once.
Compaction is aggressive: one August event removed 96.6% of context (607,378 → 20,766 tokens), and 8 of 36 pre-summary messages left no trace.
On Aug 13, a hook emitted 18,057 bytes but only 11% reached the model — runtime said success, logs said complete, selftests passed 9/9. Only byte-by-byte comparison of script output vs. transcript caught it: hook output above 10,000 units goes to a file, leaving only a 2 KB preview. The map is now split into 5 slices under the ceiling.
Since Sep 21, 64 of 65 compactions happened at or below 365K tokens (previously sessions grew to 1M).
Guards only count after rejecting a deliberately planted defect: 52 of 74 guards had that proof by Sep 25; the rest are declared debt.

The article ends with a 7-step build order, each step tied to a native feature and a "red" failure test that involves deliberately breaking the piece. Limitations: one operator, one installation; the study doesn't assess whether delivered memories were correct or useful.

Original post →

61 days of Claude Code transcripts: hooks reported success but only 11% of context reached the model

More from coding & agent