memory reads feel unreliable across sessions in my setup

wanted to flag something i've been hitting with the memory system in 4.6. could be just me or my workflow, so take this with a grain of salt.

writes are fine. files show up under .claude/projects/ where they should. the issue is what happens next session. memories don't reliably get picked up. sometimes they do, sometimes they don't, and i can't tell from the outside which is which. it seems worse when my first message of a session is task focused instead of conversational.

specific things i've seen:

• the model re-asks me stuff that's already in memory
• preferences i told it to remember don't always carry over
• there's no visible signal that memory was actually consulted

writes aren't the problem. reads are.

how i ended up working around it: i wrote a small mcp server (https://github.com/devnullnoop/MGCP) that hooks into session start and forces a context load before the first response. it also queries memories by task relevance instead of returning everything as one blob, and it splits memory into a few categories (preferences, project state, lessons, feedback) so retrieval is more targeted. that's been a lot more reliable for me, which is partly why i'm bringing this up. the capability is there, but the trigger and filtering layers feel like they're missing.

if any of this would be useful to dig into more i'm happy to share specifics.

env: claude code cli, macos 15.7.5, multiple projects over a few weeks.