feature-memory

Feature Memory (LLM-FM)

A documentation compiler for software repos. Maintains feature-level memory so AI coding agents don't rediscover the project from scratch every session.

The Problem

Every project has two codebases. The real one (files, tests, routes, configs). And the remembered one (what it does, why, which files matter, what's dead). The second one is more useful and decays fastest.

Feature Memory makes it explicit: a structured layer of feature documentation that lives alongside your code, stays current through hooks and skills, and gives agents immediate context when they start a session.

Install (Claude Code Plugin)

claude plugin marketplace add h5kk/LLM-FM
claude plugin install feature-memory@h5kk-plugins

Then open a Claude Code session in your project and say:

Initialize feature memory

The agent scaffolds the docs structure, scans your codebase, proposes features, and creates initial feature pages. No scripts to run — the plugin handles everything through skills.

How It Works

docs/feature-memory/
  index.md                    # feature table + auto-generated Mermaid relationship diagram
  recent.md                   # last 5 days of activity
  changelog.md                # append-only global log
  changelog-viewer.html       # interactive changelog browser (open in any browser)
  features/
    auth/                     # large feature — split mode
      index.md                # navigation hub
      product.md              # product audience: what it does for users
      engineering.md          # engineering audience: files, routes, patterns
    billing.md                # small feature — single file
  changelogs/
    changelog.json            # compiled changelog data (loaded by viewer)
  reports/                    # lint reports, reorg proposals

.feature-memory/
  config.yaml                 # globs, route patterns, mode (small | split | mixed)
  events.jsonl                # hook event log (current session)
  events-{session-id}.jsonl   # archived per-session logs

Three hooks wire into Claude Code's lifecycle:

Hook	When	What it does
SessionStart	Agent opens a session	Archives previous session events, injects feature list + recent activity
PostToolUse	After every Edit/Write	Logs the edit, reminds agent which feature docs to update (split-mode aware)
Stop	Session ends	Captures git author, compiles changelog.json, reports features with missing doc updates

A companion skill (feature-memory) gives the agent a structured workflow for reading and updating feature pages, changelogs, and the index.

Dual Developer + Product Changelogs

Every changelog entry is tagged with an audience:

• Developer — internal changes: refactors, dependency bumps, test changes
• Product — user-facing changes: new behavior, UI changes, API contract changes
• Both — when the change matters to both audiences (default)

The Changelog Viewer (docs/feature-memory/changelog-viewer.html) renders these as a searchable timeline:

• Tabs: All | Product | Developer
• Search by feature, author, summary, file path, or tag
• Tag chips on every entry — up to 5 auto-inferred tags per category (Impact · Quality · Process · Tech)
• Tag filter panel: click to filter; AND across categories, OR within a category
• Toolbar: Expand All, Copy JSON, Copy Markdown, Export CSV
• Git author attribution with consistent color coding
• Timeline grouped by date with expandable entries showing paths and full git details
• Works offline — data is embedded inline, no server required
• Auto-upgrades when a newer plugin version is installed (preserves all existing data)

Large Feature Support (Split Mode)

Set mode: split for a feature to get a sub-folder structure instead of a single file:

# .feature-memory/config.yaml
features:
  auth:
    mode: split        # creates features/auth/{index,product,engineering}.md
    globs:
      - "src/auth/**"

Sub-pages carry their own audience frontmatter so the Stop hook can route doc updates correctly without reading the parent file.

Relationship Diagram

docs/feature-memory/index.md includes an auto-generated Mermaid diagram compiled from [[feature-id]] relationship links across all feature pages. The maintainer skill regenerates it when relationships change.

Quick Start

1. Install the plugin (see above)
2. Open Claude Code in your project
3. Say "initialize feature memory" — the init skill walks you through setup
4. Say "update feature memory" after making code changes — the maintainer skill keeps docs current

Agent commands