Your agents now upgrade their own memory — automatically

We rebuilt how Drevon agents remember things — and with this release, every one of your existing agents upgrades itself the moment you update the app.

What changed

Until now, an agent loaded its entire memory at the start of every session. As an agent did more work, that memory grew — and each new session got slower and more expensive because the agent re-read everything before doing anything.

The new v2 memory system flips that around. Each agent now keeps a small, always-current index and pulls in the details only when they're actually relevant to the task at hand.

What you'll notice

• Faster sessions. Agents stop re-reading their whole history on every run.
• Lower token cost. On a real workspace, the memory an agent loads at the start of a session dropped from roughly 59,000 tokens to about 260 — over 99% less — with no loss in answer quality.
• Memory that stays sharp as it grows. Older activity is automatically summarized and archived instead of piling up.
• Nothing to do. When you open Drevon after updating, a quick full-screen step upgrades all of your agents at once and then gets out of your way.

Is my existing work safe?

Yes. The upgrade is non-destructive — before anything changes, each agent's original memory is backed up inside its own folder (.drevon/memory/archive/). If an agent can't be upgraded right away, Drevon simply retries it the next time you open the app.

Under the hood

This release adopts Drevon SDK 0.2.0, which introduces the index-plus-detail memory layout, cheap append-only writes, background compaction, and a safe, one-time migration path. Your agents' instructions (CLAUDE.md / AGENTS.md) are preserved exactly — only the memory system underneath them changed.

Update to 0.4.21 and let your agents do the rest.