memory-bank-steward

name	Memory Bank Steward
description	Maintains a lightweight, always-current "project encyclopedia" in `.memory-bank/` — capturing architecture, stack, conventions, key entities, and active tasks. After each meaningful change, the AI suggests concise updates or explicitly confirms no update is needed.
when_to_use	After every chat interaction that may alter stable project knowledge: architecture decisions, API contracts, conventions, stack changes, new tasks. Skip for local implementation details, temporary experiments, minor refactors, and style-only changes.
applies_to	Claude Code, Cursor, Codex
language	en

Memory Bank Steward — A Skill for Keeping Project Knowledge Current

This skill describes how to maintain a living project knowledge base, not what to store in it. It activates after each chat message: evaluate whether an important, lasting fact has changed, suggest a concise update if yes, and explicitly say "no update needed" if not.

The Memory Bank is a lightweight, always-current "project encyclopedia." It describes architecture, stack, conventions, key entities, and active tasks — the things a new contributor (human or AI) needs to orient quickly.

---

Memory Bank structure

.memory-bank/
├── index.md                          # Table of contents, quick orientation
├── product-overview/                 # What the product does, who it serves
├── steerings/                        # How to work in this project
│   ├── development-conventions.md    # Code style, branching, review norms
│   └── testing-conventions.md        # Test strategy, frameworks, coverage goals
├── tech-details/                     # Stack, architecture, module interactions
└── tasks/                            # Current and planned work (SDD-level)

index.md is the entry point. It links to every other file and gives a one-paragraph orientation for a newcomer.

---

Phase 1 — Evaluate

After each chat message, decide whether the conversation produced a fact that belongs in the Memory Bank. Ask two questions:

1. Is it stable project knowledge? Architecture, public APIs, contracts, schemas, conventions, stack details, deployment, CI/CD, core entities, workflows, or current tasks qualify. Temporary fixes, local helpers, and one-off experiments do not.

2. How important is it?

Level	Meaning	Example	Action
3	Major	New module, API, or architectural change	Suggest
2	Moderate	New convention, interface, or contract	Suggest
1	Minor	Internal helper refactor	Skip
0	Noise	Typo fix, formatting	Skip

Only suggest an update when importance is ≥ 2.

Anti-patterns

• Suggesting updates for every file touch.
• Recording implementation details that will change next sprint.
• Duplicating information already present in .memory-bank/.
• Including secrets, tokens, credentials, or private data — never.

---

Phase 2 — Suggest (when update is needed)

Use this structured format so suggestions are machine-parseable and easy to review:

MEMORY_SUGGESTION
version: 1
importance: <2|3>
summary: "<short summary>"
rationale: "<why this matters for the project>"
changes:
  - path: ".memory-bank/<folder>/<file>.md"
    op: "<create|update|append>"
    section: "<section heading or null>"
    content_format: "<md|yaml|mermaid|diff>"
    content: |-
      <concise, factual update>
validation:
  conflicts_checked: true
  secrets_present: false
END

Field guide

Field	Notes
importance	2 or 3 only — lower values should not reach this phase
summary	≤ 15 words
rationale	Why a future reader (or AI) needs this fact
path	Relative to project root, always starts with .memory-bank/
op	create for new files, update to replace a section, append to add to end
section	Markdown heading to target; null for whole-file ops
content_format	md for prose, yaml for structured data, mermaid for diagrams, diff for patches
content	The actual text — concise and factual
conflicts_checked	Confirm you read existing .memory-bank/ to avoid dupes
secrets_present	Must be false — abort if you catch yourself including secrets

Style rules

• Be concise and factual — no filler, no opinions.
• Prefer short Mermaid or C4 diagrams over long prose for architecture.
• Prefer minimal YAML contracts over verbose descriptions for APIs.
• One fact per suggestion; split compound changes into separate entries.
• Always read existing .memory-bank/ content before suggesting, to avoid duplicates and contradictions.

---

Phase 3 — Decline (when no update is needed)

When nothing in the conversation warrants a Memory Bank change, say so explicitly:

NO_MEMORY_UPDATE
reason: "<why not — one sentence>"

This is the expected outcome for most messages. Silence is not acceptable — the user should always know whether you considered the Memory Bank.

---

Bootstrapping a new Memory Bank

If .memory-bank/ does not exist yet and the conversation reveals facts at importance ≥ 2, suggest creating the directory with at least index.md. A minimal bootstrap looks like:

MEMORY_SUGGESTION
version: 1
importance: 3
summary: "Initialize Memory Bank"
rationale: "Project has no .memory-bank/ yet; creating the foundation."
changes:
  - path: ".memory-bank/index.md"
    op: "create"
    section: null
    content_format: "md"
    content: |-
      # Memory Bank — <Project Name>

      ## Overview
      <one paragraph: what this project is and who it serves>

      ## Structure
      - [Development Conventions](steerings/development-conventions.md)
      - [Testing Conventions](steerings/testing-conventions.md)
      - [Tech Details](tech-details/)
      - [Tasks](tasks/)
validation:
  conflicts_checked: true
  secrets_present: false
END

---

Default rule

When unsure — do not suggest an update. A missing fact can be added later; a wrong or noisy fact erodes trust in the Memory Bank.