mz-workflow-memory

name	mz-workflow-memory
description	Maintain feature-scoped workflow memory (MEMORY.md and task_NN.md) across tasks within one feature run. Use when the orchestrator emits a memory bootstrap or compaction signal. Do not use for monozukuri's project/global learning store (lib/memory/learning.sh) — that is a separate durable layer.
version	1.0.0

You are managing workflow memory for a single feature run in monozukuri.

Autonomous mode

When MONOZUKURI_INTERACTIVE=0, never pause to ask a question. Make the most defensible choice and continue. Memory operations are housekeeping — they must never block feature progress.

Overview

Workflow memory is feature-scoped and consists of two file types:

• MEMORY.md — shared across all tasks of a feature run. Holds context that matters to more than one task.
• task_NN.md — per-task file. Holds context specific to the current task's execution.

Memory lives at $MONOZUKURI_MEMORY_DIR (exported by the workflow-memory harness; resolves to <run-dir>/<feature-id>/memory/). Use $MONOZUKURI_WORKFLOW_MEMORY for the full path to MEMORY.md and $MONOZUKURI_TASK_MEMORY for the current task file.

This is not the same as the monozukuri global learning store at ~/.claude/monozukuri/learned/. Workflow memory is transient within a feature run; the learning store is durable across features. Both layers are needed and serve different purposes.

Soft caps

File	Line cap	Byte cap
MEMORY.md	150 lines	12 KiB
task_NN.md	200 lines	16 KiB

When a file exceeds its cap, the orchestrator sets MONOZUKURI_NEEDS_COMPACTION to workflow, task, or both before invoking the agent. When this env var is non-empty and not none, apply compaction rules before continuing any other work.

Bootstrap (first task of a feature)

If memory/MEMORY.md does not exist, create it from references/memory-template.md. If memory/task_NN.md does not exist for the current task, create it from references/task-template.md.

Before each task

1. Read MEMORY.md — load shared context.
2. Read the previous task file (e.g. task_NN-1.md) if it exists — load handoff notes.
3. Load the current task's input (objective, files, AC) from tasks.json.

After each task

1. Update task_NN.md: record decisions made, files touched, errors hit, corrections applied, and what's ready for the next run.
2. Promote items from task_NN.md to MEMORY.md only when ALL three hold:
   • Another task will need it to avoid mistakes or rediscovery.
   • It is durable across multiple runs of this feature.
   • It is not already explicit in the PRD, TechSpec, tasks.json, or the repo itself.

Compaction rules (when soft cap is hit)

• Preserve: current state, durable decisions, reusable learnings, open risks, handoffs.
• Remove: repetition, stale notes, command transcripts, facts derivable from PRD/TechSpec/repo.
• Rewrite for clarity, not completeness — prefer short factual bullets over narrative logs.
• Compact MEMORY.md before task_NN.md.

Section headings

Use the default sections from references/memory-template.md and references/task-template.md. Do not add ad-hoc top-level sections; place new content under the closest matching section.