Kandela
Beta · Invite only

Long-term memory
for Claude Code.

Claude forgets everything between sessions.
Kandela remembers the decisions, gotchas, and context that matter.

3-seed A/B test
90 sessions · 2 scenarios

MedFlow (HIPAA pipeline) + InfraBot (Slack infra bot). Claude Sonnet 4.6, automated end-to-end.

-79.5pp
Decision Mistakes
Avg across MedFlow (−88.1pp) & InfraBot (−70.8pp)
-37%
Code Bloat
2,152 vs 3,441 lines — wrong code never written
-21%
Time in Trap Sessions
No time wasted implementing wrong decisions

Full methodology →

How it works

🧠

Remembers decisions

Why did you choose Naver Maps over Google Maps? Why is the cache TTL 2 hours? Kandela stores the why, not just the what.

Instant recall

At session start, Kandela automatically loads relevant memories — gotchas, past decisions, infrastructure notes — so Claude starts informed.

🛡️

Guards against drift

When Claude suggests changing something you decided against, Kandela surfaces the original reasoning before you regret it.

Get started in 3 steps

1

Sign up for beta

Request access at [email protected]. Beta is free. No invite code? Just email us.

2

Create an API key

Dashboard → AccountGenerate API Key. Takes 10 seconds.

3

Run the installer

curl -sL https://kandela.ai/install | bash

Enter your API key when prompted. Done in ~60 seconds.

From the blog

Deep dives into AI memory, Claude Code internals, and the science behind Kandela.

Browse all posts →