Most AI tools summarize. We don't. Here's why that matters.
The lossy middle
When Claude helps you debug for an hour and you close the tab, three things vanish: the dead ends you ruled out, the exact phrasing that made it click, and the small decisions that only make sense in context. Summaries keep the headline and drop the texture.
If you've ever asked "wait, why did we rule out Postgres again?" six weeks later,
you've hit the lossy middle. The summary your tool kept said "decided on Neo4j for
graph queries." The reasoning — the three Postgres approaches you tried, the
specific EXPLAIN output that killed the second one, the Discord conversation
that confirmed the trade-off — is gone.
What Memo keeps
- Every message, verbatim.
- Every tool call Claude made, with arguments.
- Every file it touched, with the exact diff.
- Every error, retry, and recovery.
Not summarized. Not condensed. The whole session, stored as it happened.
Why that's not expensive
The intuition most people have is "storing everything must be prohibitive." It isn't, for a reason that's worth stating plainly: a 90-minute coding session, including every tool call and diff, compresses to roughly 40 KB of text. A year of daily sessions is ~15 MB. You can store a lifetime of your AI work for less than one phone photo per day.
The hard part isn't storage. It's retrieval — making the right 2 KB surface when you need them, out of a growing pile of megabytes. That's what the next post is about.