Cursorkleosr

cursorkleosr Agent memory that cannot be forgotten — one always-on Cursor rule plus root markdown state. Skills wait to be invoked. Rules load every session.The agent reads your project state before it thinks — no more starting from zero.

---

The Breakthrough

Most agent-memory setups are opt-in: a skill the agent must remember to call, a file the human must paste into chat. This one is always-on.

.cursor/rules/workflow-memory.mdc ships with alwaysApply: true, so Cursor injects the operating loop into every agent session automatically. The agent's first move is reading workflow_state.md — before you ask for anything. When it finishes meaningful work, it writes state back. Memory stops being a feature you invoke and becomes a property of the repository.

Setup

git clone https://github.com/kleosr/cursorkleosr.git
cd cursorkleosr

Open project_config.md, set your goals and stack. Open workflow_state.md, set Phase: INIT — READY. Done — no dependencies, no build step.

Usage

There is nothing to invoke. Open the repo in Cursor and work:

• Session start — the agent reads workflow_state.md and project_config.md, then reconstructs the next action on its own.
• During work — one phase at a time: ANALYZE → PREPARE → IMPLEMENT → VALIDATE.
• After meaningful change — the agent writes state back. No per-message noise.

Manual fallback, from any terminal:

cat workflow_state.md   # where are we?
cat project_config.md   # what are we building?

Files

File	Purpose
AGENTS.md	Single source of truth agent handbook: Rules, Skills, Workflows, Memory
.cursor/rules/workflow-memory.mdc	Thin adapter importing @AGENTS.md for Cursor
workflow_state.md	Moving parts: current phase, plan, blockers, log
project_config.md	Slow-changing facts: goals, stack, patterns, constraints, changelog
Instructions.md	Human-facing reminder of the read → act → write loop

Workflow Loop

read state → read config → act → write state back
                     ↓
            ANALYZE → PREPARE → IMPLEMENT → VALIDATE

graph LR
    S[Session Start] --> A[Read workflow_state.md];
    A --> B{Task Ready?};
    B -- Yes --> C[Process Next Task];
    B -- No --> D[Continue Current Work];
    D --> E[Execute Step];
    E --> F[Update workflow_state.md];
    F --> G{Validation Needed?};
    G -- Yes --> H[Run Checks];
    G -- No --> D;
    C --> I{More Tasks?};
    I -- Yes --> J[Reset Workspace];
    J --> K[Fresh Start];
    K --> A;
    I -- No --> L[Mark Complete];
    L --> M[Idle];

Architecture

.
├── .cursor/
│   └── rules/
│       └── workflow-memory.mdc  — thin adapter importing @AGENTS.md
├── AGENTS.md                    — canonical agent handbook (Rules, Skills, Workflows, Memory)
├── project_config.md            — goals, stack, patterns, constraints, changelog
├── workflow_state.md            — phase, plan, blockers, log
├── Instructions.md              — human loop reminder
└── LICENSE                      — MIT

Zero dependencies. No build step. Root markdown stays editable in any editor, diffable in any tool, and greppable from any terminal.

Development

Documentation-only project. No tests, no compilation, no CI pipeline. Validate edits with:

git diff --check

Why Markdown?