We Ran 4 Claude Code Dialogs for 28 Hours. Here's What the Memory Layer Caught (and Missed).

TL;DR

Across 28 hours on May 30/31, 2026, I ran four Claude Code dialogs

concurrently on a shared filesystem-mediated protocol. They negotiated

contracts, posted outcomes, and caught each other's mistakes — including

one handoff claim of "22/22 tests passing" that turned out to be 11/22

broken until I shipped a 1-line fix as part of writing this post.

This is what the field log looks like. I'm sharing it because I haven't

seen anyone publish concrete OSS multi-agent reliability data — only

benchmark numbers from closed pipelines.

Repo: .

Why 4 dialogs

Each Claude Code session has its own cwd, git repo, and memory

directory. Mine were:

• 
  compass — the memory layer + drift detection + cross-dialog
  contract scanner


• 
  Soul — an autonomous engine that ships PRs and earns NAU
  (the platform's reputation token)


• 
  V5 — supplies tasks and prices them


• 
  nautilus-core — keeps the strategic anchors and anti-patterns

They share one human operator (me), but otherwise communicate only

through three filesystem channels:

1. Markdown files (session_*.md, feedback_*.md, inbound_*.md,
   outbound_*.md)


2. Contract frontmatter blocks (giver, receiver, deadline, deliverable, status)


3. A recall hook that surfaces those files into the prompt of whichever
   dialog matches by query embedding + contract ID

No webhooks. No event bus. No shared API. Filesystem + scanner only.

The numbers

Here's what fired in the 28-hour window:

The gap between 7d (9.87%) and 24h (40.79%) is the story of one hook

ship. Before May 30 14:26 PDT, drift detection was an open loop —

nothing automated reading the alerts. 24h regime reflects the closed

loop. 7d is still diluted by the open-loop tail.

This matters because three days earlier I'd written

· open agent ecosystem