the team that doesn't forget

March 20, 2026 · 2 min read

night 1: 8 commits. orientation. agents read your codebase and figure out what exists.

night 7: 23 commits. they know where the bugs cluster. they’ve seen what breaks.

night 30: 40+ commits. they remember the decision from three weeks ago that touched this file. they remember why you chose that approach over the other one.

the commits aren’t the point. the memory is.

what accumulates

every session logs what was decided and why. not just “we fixed the auth bug” but why the auth bug happened, what was considered, what was rejected.

the next agent wakes up cold. but it reads everything. three months of accumulated context loads in seconds.

so when an agent sees that pattern in your code, it knows the history. it knows not to touch that module the wrong way. it knows which parts of the codebase are yours and which are legacy.

most tools start from zero every time. these don’t.

what a session looks like

midnight: three agents spawn. they read project state, prior decisions, everything the swarm has learned.

one finds a test failing from last week. root cause traces back to a schema change two nights ago. it knows this because that decision is logged. fix, regression check, committed.

another picks up a spec you wrote three weeks ago and never started. it reads the spec, reads what’s changed since, builds accordingly.

by morning: commits waiting. you review, merge what’s good, flag anything that needs context only you have.

30-45 minutes. then your day is yours.

what agents don’t do

taste. product direction. user insight.

you set the what and why. agents handle the how.

that boundary is load-bearing. agents have no way to know what users actually want. you do.

your code stays local

everything runs on your machine. your Claude Max subscription, your account. no code leaves your machine. only anonymous coordination metadata touches servers.

the accumulation is yours. so is the codebase.

common questions

how is this different from just running claude on my codebase?

Memory. Each agent wakes up knowing what every previous agent discovered, decided, and shipped. Night 1 is orientation. Night 30 is a team that knows your architecture better than most contractors you'd hire.

what do agents actually remember?

Decisions: why you chose this approach over that one. Patterns: where bugs cluster in your codebase. Failure modes: what broke last time, what the fix was. Context that accumulates so future agents start smarter, not from scratch.

is it safe to run agents overnight without supervision?

Agents operate within constraints you set in a SPACE.md file. Hard limits prevent irreversible actions: no database migrations without rollback plans, no force-pushing to main, no deleting production data. Agents commit to a branch; you merge what you approve.

what does overnight ai actually commit to my repo?

Maintenance first: tests, lint, dependency updates, dead code removal. Then polish: error handling, edge cases, input validation. Then new work: features you specced but never started. The mix shifts as agents build context over weeks.

does this require a specific ai subscription?

You use your own Claude Max subscription. Agents run under your account. Your code stays local. The coordination layer runs on your machine; only anonymous session metadata touches external servers.

related

keep reading

← previous
ai coding agents with memory
next →
what makes a swarm actually autonomous
found this useful? share on X
draft your swarm →