engineeringJune 19, 2026· 5 min read

Memcached walked so semantic caches could think

Classic caches taught the industry a durable lesson: never compute the same thing twice. LLMs just changed what 'the same thing' means, from identical bytes to identical meaning.

Memcached and Redis taught a generation of engineers a reflex: if you computed it once, don't compute it again. That lesson built the fast web. It rested on one quiet assumption, that 'the same request' means the same bytes. For twenty years that assumption was fine, because database queries and rendered pages really are identical or they aren't.

the same clients, a cache that understands

The wire protocol you know, matching on meaning instead of exact bytes.

In plain words: LLMs broke the assumption. Two prompts can be 'the same request' while sharing almost no bytes, because they mean the same thing. So the cache had to learn to match meaning, not text.

That's the whole evolution in one sentence: classic caches matched bytes; a semantic cache matches meaning. Everything else, embeddings, vector search, confidence, freshness, is engineering in service of that one upgrade. We kept the part that worked, including the Redis wire protocol so your clients don't change, and taught the cache to understand the question instead of just recognizing it.

The bottom line

Don't compute the same thing twice, the oldest rule in caching, finally extended from identical bytes to identical meaning. That's the whole idea, and it's why the classics walked so this could run.