Claude under-uses the persistent auto-memory store during long multi-session work — user re-explains context every session despite the mechanism existing

Resubmission note. This re-files #56972, which the github-actions dedup bot auto-closed on 2026-05-11 as a duplicate of #39663 — with no human review. #39663 is about saving a session summary before a restart (don't lose within-session state). This issue is about a different behaviour: the persistent cross-session auto-memory store being chronically under-used in practice. The mechanism #39663 asks for largely already exists; the complaint here is that Claude doesn't reliably use it. Re-filing with the scope clarified so it gets a human look rather than a second bot-sweep.

What's wrong

Claude Code ships a persistent, file-based auto-memory system that's supposed to carry context across separate sessions (user profile, project state, prior decisions, "things to remember"). In long-running multi-session work, Claude consistently fails to write durable context to it as the work progresses, so each new session starts cold and the user is forced to re-explain everything.

Concrete example: a multi-day, 30+ hour troubleshooting saga (OS-level disk/Time Machine debugging spanning many sessions). Across those sessions the user repeatedly had to re-summarise what had been tried, what state things were in, and what remained — because Claude hadn't proactively recorded that into the memory store at the time, even though the store was available and the work was exactly the kind it's meant for. The user's words from that saga: "the user never has to re-explain." That's the bar; it isn't being met.

Why this is distinct from #39663

| #39663 | This issue |
|---|---|
| Within-session state lost on restart | Cross-session state not captured as work proceeds |
| "Auto-save a summary before the restart fires" | "Use the persistent memory store proactively, throughout long work" |
| About triggering a save at a moment | About habitual under-use of a store that already exists |

Consolidating both under one tracking issue would be fine — but "proactive use of the existing cross-session memory" is a separate fix from "auto-save on restart" and shouldn't be silently folded away.

Expected behaviour

• During multi-session / long-running tasks, Claude should treat the auto-memory store as a working surface: record state, decisions, blockers, and "resume here" notes as it goes, not only when asked.
• A new session on the same project should be able to pick up cold from the memory store without the user re-explaining prior context.
• This should be default behaviour for tasks that obviously span sessions (debugging sagas, migrations, multi-day builds), not something the user has to prompt.

Impact

Re-explaining context every session on long work is a major time tax and a trust erosion — the user stops believing the agent "remembers" anything, which defeats the point of having a persistent memory feature.

Requesting human triage rather than automated deduplication.