Kairos in Claude Code

Kairos is the codename of an unreleased feature found in the source code of Claude Code, Anthropic's AI assistant for programmers, which accidentally became public at the end of March 2026. According to that code, Kairos would run in the background and tidy up the assistant's memory while the user is away. In the published version it was switched off, so no user could turn it on.

Where does the name come from?

On 30 March 2026 Anthropic published version 2.1.88 of Claude Code on npm with a source map, a debugging file that pointed to the full source code. More than 512,000 lines of code became public, and developers found several features that were built but hidden (The Batch, 3 April 2026). One of them was a subsystem called Kairos, Greek for "timely" according to The Batch. The blog post on what the leaked source revealed covers the whole leak.

What would Kairos do?

The Batch describes Kairos as a subsystem that "would run as an always-on background agent". Its logic, called autoDream, merges duplicate memories, eliminates contradictions, resolves speculations and prunes what is stored, so that the memory is more useful for the next action. InfoQ calls it an autonomous daemon mode that consolidates memory while the user is idle. Developer Alex Kim, who read the code on 31 March, found code paths for:

• a /dream skill for "nightly memory distillation";
• daily logs that are only ever added to;
• subscriptions to GitHub webhooks;
• background daemon workers;
• a scheduled refresh every five minutes.

He added that the implementation was heavily gated, "so who knows how far along it is".

Was it released?

No. According to The Batch, these features sat behind flags that compile to "false" in the published build, a sign that they were still in progress. Up to version 2.1.281 of 23 September 2026, Claude Code's changelog mentions neither Kairos nor a /dream command. A comparable memory clean-up did appear in another product: on 6 May 2026 Anthropic released Dreams as a research preview in Claude Managed Agents.

Why does an agent need a memory clean-up?

Memory only helps an agent if it is current. Anthropic's Dreams documentation describes the problem: an agent writes to its memory as it works, and over many sessions the store accumulates duplicates, contradictions and stale entries. Kairos would have cleaned up automatically inside Claude Code. Dreams runs the clean-up as a separate job that reads the memory store and past sessions and writes a new store. The input store is never modified, so you can review the result and discard it.

Example

Suppose a project's memory holds two notes: an early one saying the app stores its data in PostgreSQL, and a later one saying it moved to SQLite. An autoDream-style clean-up would keep the later fact and drop the contradicted one, so the next session does not start from the old database. According to its documentation, Dreams also replaces stale or contradicted entries with the most recent value.