Upgrading a real shared-memory workflow to MemPalace v3.3.0: what actually improved

[fuzzymoomoo]

We started using MemPalace very early, and it quickly became continuity
infrastructure for our real workflow: handoff capsules, shared-memory cold
starts, and multi-project engineering memory across Hades, Anubis, CDD, and
related repos.

Because of that, I did not want to treat v3.3.0 as a blind upgrade. I
compared our older working 3.1-era local setup against an isolated v3.3.0
candidate on the same machine, using:

• the shipped synthetic benchmark suite
• a small workflow-local golden query pack
• a copied real palace snapshot

The honest result is: this was worth it, but mostly as an operational upgrade
rather than a dramatic ranking upgrade.

The clearest benchmark win was concurrent search behavior:

• concurrent avg latency: 116.6 ms -> 80.4 ms (-31.0%)
• concurrent p95: 233.5 ms -> 128.8 ms (-44.8%)
• concurrent p99: 287.4 ms -> 172.6 ms (-39.9%)

That matters for us because MemPalace is not just a note store in this setup.
It is part of the shared-memory path between sessions and between agents.

The strongest functional win was migration and recoverability. Our copied live
palace initially looked unreadable from the newer runtime, but migrate could
explain the source format, recover the data via SQLite, and rebuild it cleanly.
After fixing one Windows-specific handle-release issue in the migrate swap
path, the copied palace migrated successfully and reported the full drawer
count correctly.

That status point mattered too: the old working setup surfaced 10000
drawers, while the migrated candidate correctly reported 16468.

What did not improve clearly in this run was workflow-local CLI ranking. On
our current golden query pack:

• top-1 accuracy stayed 0.2 -> 0.2
• top-5 recall stayed 0.6 -> 0.6
• average CLI latency was basically flat: 1505.99 ms -> 1488.86 ms

So I would not frame v3.3.0 as "search quality got dramatically better"
based on this dataset.

I would frame it as a meaningful upgrade for power users because it improved:

• migration visibility and recovery
• truthful large-palace status
• concurrent search behavior
• post-migration consistency on a copied real palace

We also turned the Windows migrate swap fix into an upstream draft PR so that
part is no longer just a local workaround:

• MemPalace/mempalace#890

My takeaway: if you are still on an early 3.x setup and using MemPalace as
real continuity infrastructure, v3.3.0 is worth evaluating. Just do it with
a copied palace, a saved baseline, and a small golden query pack so you can
measure your own workflow honestly.