# 60+ compactions into one Claude Code session, the summary had kept 0 of 12 older decisions. How do you keep yours? — retained record

> Retained reddit record from reddit:ClaudeCode.

Canonical: https://highsignal.app/data/records/15c9193a380522e4

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:ClaudeCode)  published 2026-10-21 06:26 ingested 2026-10-22 02:42

[original source ↗](https://www.reddit.com/r/ClaudeCode/comments/1xbfly6/60_compactions_into_one_claude_code_session_the/)

## retained text



```
*\[Edit\]*  
  
A few replies asked **why I stay in one session**, so here's the context.

I use Claude Code for more than code. I think through product strategy with it, using our own skills to pressure-test ideas, and then fold the result straight into the codebase in the same session. Over weeks I kept updating the rules and wrote down how Claude answers and pushes back when it checks my claims.

I used to do the usual thing: when the context hit around 80%, I'd have Claude write a handoff doc and move to a fresh session. That doc wasn't just facts. I put the way of working into it in a lot of detail: how to check my reasoning, how to answer, what to push back on, along with the rules and skills. It still didn't work. The new session knew the facts but didn't work that way. I tried for a while, then gave up and stayed in one long session, letting it compact. When something old is needed, Claude greps the transcript.

So I'm stuck between two bad options.   
A fresh session loses the way of working, even when it's written down in detail. A long session keeps the way of working but slowly loses old decisions, as below. Has anyone gotten a new session to actually work the way a written handoff describes?  


\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_

  
Disclosure first: I'm building a local AI agent called 'Sonjari' :)

Setup: Claude Code desktop, Opus 5.5. One long session since Sep 30 on a Windows agent project, with [CLAUDE.md](http://CLAUDE.md) and auto memory on. It has been compacted 62 times so far.

What I expected: the compaction summary would keep the decisions we made along the way.

**What I found when I checked:**

\- I took a stretch from about 10 compactions back and listed 12 decisions we made there.

\- The current summary had 0 of the 12.

\- Our memory files had about 8. The other 4 were nowhere.

\- The most recent stretch was fine: 20 of 21 items were in the summary.

Recent work survives. It's the summary of a summary of a summary that quietly drops the older stuff.

The 4 we lost completely were not things we could do without: a one-line product positioning we'd agreed on, a design call to do some cleanup in one place on the backend, a bug we had tracked down, and how to answer one specific comment. I only noticed because I asked "what did we decide about X?" and Claude had to piece the answer back together from the summary.

>*How do you handle this? A decision log you update by hand? A hook before /compact? Or do you just start fresh sessions more often?*

One thing we're trying: while 'Sonjari' is idle, it reads the Claude Code transcript on my own GPU and keeps decision cards, each with the line number and a word-for-word quote, so Claude can pull them back after a compaction. Sonjari itself is out, but this feature ships in the next update, and I haven't measured whether it actually helps. Curious what works for you first.
```



document fetched 2026-10-22 00:43 · canonical https://reddit.com/r/ClaudeCode/comments/1xbfly6/60_compactions_into_one_claude_code_session_the
