Two Sides of the Same Coin: The Stalactite Principle and autoDream, Five Months Apart
ℹ️
Disclaimer: I am not affiliated with, employed by, or sponsored by Anthropic. I am a paying Claude Pro subscriber and receive no compensation or endorsement for mentioning Claude in this article. All opinions and experiences shared are my own.
On November 4, 2025, I documented something I called the Stalactite Principle. The core idea was simple: AI coding sessions are ephemeral, but knowledge shouldn't be. By externalising context to GitHub Issues — letting insights accumulate like mineral deposits over time — you could give Claude Code a persistent memory that survived terminal restarts, context resets, and session boundaries.
I built it out of laziness. Re-explaining the same architectural decisions every session felt like a tax on thinking. The Issues became the externalisation layer. The model became stateless. The knowledge stayed.
Five months later, on March 31, 2026, a packaging error in @anthropic-ai/claude-code v2.1.88 accidentally exposed the tool's entire TypeScript source. Researchers digging through 512,000 lines found something called autoDream — a background memory consolidation system. It activates on a three-gate trigger: 24 hours elapsed, 5 sessions completed, a consolidation lock preventing concurrent runs. All three must pass. Then it moves through four phases: Orient → Gather → Consolidate → Prune.
Anthropic confirmed the incident as "a release packaging issue caused by human error, not a security breach." They said nothing about the architecture.
Two Faces
autoDream is the inside of the coin. The Stalactite Principle is the outside. You only ever see one face at a time — but they're the same metal.
Both are solving the same problem: AI agents lose context, and that loss is expensive. Both treat memory as something to be actively maintained rather than passively accumulated. Both separate what the model is doing right now from what the model needs to remember across time.
The difference is infrastructure. autoDream is a daemon. The Stalactite Principle is a discipline.
One runs automatically, triggered by elapsed time and session count. The other requires intent — a developer who noticed the cost of repetition and decided to externalise rather than re-explain. One is engineered. The other is improvised. Both arrive at the same architectural shape: a persistent layer, outside the model, that accumulates signal and prunes noise.
The Convergence
I didn't arrive here by reading Anthropic's internal documentation. I arrived here by hitting the same wall their engineers hit, and building around it with the tools I had — a shell function, GitHub Issues, and a .claude folder.
This post is not a claim that I built what Anthropic built. It is an observation that two independent paths, taken at roughly the same time, arrived at strikingly similar conclusions.
That convergence is itself the signal: the problem is real, the solution space is constrained, and the architectural shape of the answer is probably not a coincidence.
The Stalactite Principle is documented. The timestamp is November 4, 2025. autoDream is now public, whether Anthropic intended it to be or not.
Draw your own conclusions.
This post was written in dialogue — with Claude Chat for framing, with Claude Code for fact-checking, and with the sources themselves for honesty. Three passes to get it right. That's the methodology too.