AN AGENT ENVIRONMENT FOR CLAUDE CODE, MODELED ON HUMAN MEMORY
Claude Code runs on conversation, and conversation runs on your attention. claude‑automata drives itself: hand it months of work and it finishes in days. All the loop asks for is your intent.
The moment the agent declares itself done, a hook blocks the stop and summons the advisor: an independent metacognition with access to the whole story.
This is what ploop provides. The authority to end the loop rests with the advisor.
Saying "done" doesn't end it. Being done does.
ploop defines the intent you hand it as the anchor, plants it outside the transcript, and the advisor consults it constantly. Injected into context right after each auto‑compaction, the anchor remains the single source of truth.
Once launched, the loop does not stop, and it treats you as a tool in service of the anchor. You can drop in instructions mid‑flight or ask a separate docent session for an explanation. You are one component of the loop.
Memory that fails the verified transaction circuit is forgotten along with the loop. Forgetting is a core component of intelligence.
The human memory system, modularized into three plugins. Working memory consolidates into long‑term memory, and long‑term memory is rebuilt on a cycle.
Working memory forms inside ploop. The anchor sets the criteria, and the advisor provides the metacognition.
tx is the screening gate where working memory consolidates into long‑term memory. It checks the implementation against the intent, passes CI, and writes long‑term memory as one squash merge.
All non‑executable, git‑tracked text is long‑term memory, organized with precision: specs as semantic memory, archived changes as episodic, the glossary as lexical, skills as procedural.
Stagnant memory does not survive. The refine workflows clear out dead memory and converge the whole repository toward its optimal state.
Lets a single session sail autonomously for days. Your intent becomes the anchor that holds the loop steady against drift.
A Git workflow that treats change as a transaction (an integrity boundary). It guards the repository's integrity. Install it and the agents drive it on their own.
Large‑scale workflows that eliminate technical debt:
/refine:code optimizes the code architecture,
/refine:docs aligns documentation with the code, and
/refine:integrity audits integrity. They sweep the whole
repository, so a run can take ten hours or more.
Run one command inside a git repository and setup is complete.
REQUIRES · Claude Code · uv · POSIX (macOS / Linux / WSL)
uvx claude-automata initOne command sets up the environment claude‑automata needs to run reliably.
| permissions.defaultMode | "bypassPermissions" |
|---|---|
| model | "opus[1m]" |
| alwaysThinkingEnabled | true |
| autoCompactEnabled | true |
| autoMemoryEnabled | false |
| marketplace + plugins | registers clomia/claude-automata, enables its plugins |
| external CLIs | repomix, gh |
This is the whole operating surface. Write an anchor, launch it, and step in only when you want to.
1 · WRITE THE ANCHOR
# an agent interviews you and writes the anchor
/ploop:define-mission # a mission with an end
/ploop:define-purpose # an open-ended purpose2 · LAUNCH
# in a fresh session: hand the anchor to the loop
/ploop:launch [anchor text]3 · OPERATE
/ploop:off # pause the loop
/ploop:on # resume / restore the loop
/ploop:docent # reports the loop's progress (must run in a separate session, same directory)4 · KEEP THE REPOSITORY LEAN
# between missions: hours-long maintenance passes
/refine:code # optimizes the code architecture
/refine:docs # aligns documentation with the code
/refine:integrity # verifies logical integrityAPPENDIX · WHY AGENTS BRANCH
To guard repository integrity, tx blocks writes to the base branch.
Agents work on a tx-* branch cut by /tx:open,
and /tx:close lands it on base as one verified, CI‑green
squash merge. An agent branching off instead of editing the base branch
is tx working as intended.