I gave my coding agent memory with no embeddings, no vector DB, and no cloud

The default recipe for "agent memory" in 2026 is: embed everything, stuff it in a vector database, do a similarity search, pray the right chunk comes back. It works — and for a coding agent on your own machine, it's the wrong tool.

I built Corvusmem on the opposite bet: deterministic, offline, zero-dependency. No embeddings, no vector DB, no network. Here's the reasoning.

1. Determinism beats "probably relevant"

An embedding search gives you probably the right memory. Run it twice, get different neighbours as the index shifts. For memory you rely on — "which database?", "what's the deploy rule?" — you want the same answer every time. Corvusmem is a deterministic engine: same query, same facts, always. Easy to test, easy to trust.

2. Your memory shouldn't need a server

A vector DB is infrastructure: something to run, secure, back up, and pay for. Corvusmem's memory is a plain file in your repo. Commit it and your teammates have memory the moment they clone. Nothing phones home. No API key. No telemetry. In a world where your "memories" usually live in someone else's cloud behind an API, yours stays yours.

3. Context windows are the real budget

Embedding stores love to inject big blobs "just in case". That's tokens — every single message. Corvusmem injects a hard-capped handful of facts per message and doesn't grow as your memory grows. The cost to your context window stays flat.

4. Language-agnostic by construction

Because it isn't leaning on an English-trained embedding model, keys and values work in any language — you can store a fact in Spanish, recall it from an English prompt, and the two-hop recall still links related facts.

What you give up (honestly)

Pure fuzzy semantic sprawl. If your use case is "dump 10,000 PDFs and ask vague questions", you want retrieval + embeddings. Corvusmem is for durable, structured project memory — decisions, rules, stack, gotchas, goals — the stuff a coding agent needs to not repeat itself. Different job.

Try it

npx corvusmem install
npx corvusmem --save --dimension decision.auth \
  --value "Auth is session cookies, not JWT (CSRF handled, simpler revocation)" \
  --keys "auth,login"
npx corvusmem --recall "login"

→ How it works: https://corvusmem.vercel.app/docs

It's open, free for non-commercial use, and new enough that I'd genuinely like to hear where this bet holds and where it doesn't.