Cf Remember

$cf-remember

Extract and save project knowledge. User input: $ARGUMENTS

Purpose

After a coding session, key knowledge gets lost — logic flows, conventions, decisions, gotchas. This skill captures that knowledge into the project's memory folder so future sessions (and humans) can quickly understand the project.

Folder

Output goes to {docsDir}/memory/ (default: docs/memory/). Check .coding-friend/config.json for custom docsDir if it exists.

IMPORTANT — path resolution:

• Use MAIN_REPO_ROOT from the SessionStart bootstrap context (injected via session-init.sh). If absent, fall back to running pwd for $CWD and use $CWD as MAIN_REPO_ROOT.
• Read config from CF_CONFIG_FILE (= $MAIN_REPO_ROOT/.coding-friend/config.json) — do NOT search sub-folders
• Use CF_DOCS_ROOT as the docs base dir (= $MAIN_REPO_ROOT/{docsDir} where docsDir comes from config, default docs)
• Always resolve file_path as an absolute path: {CF_DOCS_ROOT}/memory/{category}/{name}.md
• Never use relative paths in write specs — they may resolve incorrectly when the working directory contains nested git repos

Workflow

Step 0: Custom Guide

bash "${PLUGIN_ROOT}/lib/load-custom-guide.sh" cf-remember

If the block above printed anything, apply only the ## Before, ## Rules, and ## After sections; if it shows the raw command instead of output, re-run that exact load-custom-guide.sh fence now.

Step 1: Analyze the Conversation

If $ARGUMENTS is provided, use it as a filter — focus only on conversation content that relates to the user's stated purpose. Otherwise, scan the full conversation.

Identify:

• Features worked on: What was built or modified?
• Logic flows: How does the feature work? Key decision points?
• Conventions: Patterns established (naming, structure, API design)
• Decisions: Why was X chosen over Y?
• Gotchas: Tricky parts, common mistakes, non-obvious behavior

Step 2: Determine the Category

Choose the right location based on what the knowledge is about, not just what task was performed:

Backward compat: When reading or updating existing memory files, both YYYY-MM-DD-<name>.md (new) and <name>.md (legacy) formats are valid — do not rename existing files.

Category selection guide:

• Fixing a bug → bugs/ (root cause, fix, prevention), NOT features/
• Adding/changing a feature → features/
• Establishing a pattern or rule → conventions/
• Choosing between approaches → decisions/
• Build/deploy/tooling knowledge → infrastructure/

Step 3: Assess Complexity

Before delegating to the cf-writer agent, assess the complexity of the content:

Use cf-writer agent (low reasoning effort) when:

• Simple feature docs, naming conventions, straightforward decisions
• Short content with clear structure

Use cf-writer-deep agent (medium reasoning effort) when:

• Complex architecture decisions with nuanced trade-offs
• Deep technical explanations requiring careful reasoning
• Long context that needs synthesis

Step 4: Delegate to cf-writer Agent

Dispatch cf-writer or cf-writer-deep (per Step 3) with the complete write spec as the prompt.

Check if the target file already exists:

• File exists → task: update
• File doesn't exist → task: create

Read the language setting from config (local .coding-friend/config.json overrides global ~/.coding-friend/config.json, default: en).

Build the write spec (use absolute path for file_path):

WRITE SPEC
----------
task: create | update
file_path: {CF_DOCS_ROOT}/memory/{category}/YYYY-MM-DD-{name}.md
language: {language from config}
content: |
  ---
  title: "<Title>"
  description: "<One-line summary for grep-based recall, under 100 chars>"
  tags: [tag1, tag2, tag3]
  created: YYYY-MM-DD
  updated: YYYY-MM-DD
  type: "<type based on category>"
  importance: 3
  source: conversation
  ---

  # <Title>

  ## Overview
  <1-2 sentences describing this feature/convention/decision>

  ## Key Points
  - <point 1>
  - <point 2>

  ## Details
  <Longer explanation with code examples if needed>

  ## Gotchas
  - <gotcha 1>
  - <gotcha 2>

  ## Related
  - <link to related code/docs>
readme_update: false
auto_commit: false
existing_file_action: append

Frontmatter rules:

• description is the most important field — it enables fast grep-based recall without reading full docs. Write it as a factual, searchable summary.
• Good: "Config management via .coding-friend/config.json with local/global override" — specific, grep-friendly
• Bad: "How we manage config" — too vague, won't match grep queries
• tags: 3-5 relevant keywords as array
• When task: update, update the updated date in the existing frontmatter. Do NOT change created.

Dispatch cf-writer or cf-writer-deep (per Step 3) with the complete write spec as the prompt.

Step 5: Index in CF Memory (MANDATORY)

This step is REQUIRED — do NOT skip it.

After the cf-writer agent completes and the markdown file is saved, you MUST call the memory_store MCP tool to index the memory in the database. This is a separate action from writing the file — the cf-writer agent does NOT do this.

If creating a new memory — call memory_store with:

• title: from the frontmatter title
• description: from the frontmatter description
• type: infer from category (features → fact, conventions → preference, decisions → context, bugs → episode, infrastructure → procedure)
• tags: from the frontmatter tags
• content: the full markdown content (including frontmatter)
• importance: 3 (default)
• source: "conversation"
• index_only: true
• sync_to_claude_md: true — if the memory contains project-wide rules (see below)

If updating an existing memory — call memory_update with:

• id: the memory ID (e.g., features/auth-module — derived from {category}/{name})
• content: the updated full markdown content
• tags: updated tags array (if changed)
• sync_to_claude_md: true — if the memory contains project-wide rules (see below)

AGENTS.md sync rules:

Convention memories (type: preference) are always synced to the project's AGENTS.md automatically. For other categories, set sync_to_claude_md: true when the memory contains project-wide rules, conventions, or decisions that future sessions must follow. Examples:

• decisions/api-versioning with rule "Always use URL-based versioning" → sync_to_claude_md: true
• infrastructure/deploy-checklist with rule "Run migrations before deploying" → sync_to_claude_md: true
• bugs/null-pointer-fix describing a one-time bug fix → do NOT sync
• features/auth-flow describing how auth works → do NOT sync (unless it contains rules to follow)

Rule of thumb: If the knowledge is prescriptive ("always do X", "never do Y", "use X pattern") and applies project-wide, sync it. If it's descriptive (how something works, what happened), don't.

If the MCP tools are unavailable, log a warning to the user but do NOT fail silently — the user should know the memory was saved as a file but NOT indexed.

Step 6: Confirm

Show the user a 2-line summary:

• Markdown file: path/to/file.md (created or updated)
• Memory DB: indexed ✓ — or: MCP unavailable, file only

Interpreting $ARGUMENTS

$ARGUMENTS is free-form user input. It can express:

1. A focus/filter (most common) — the main idea, purpose, or requirement that narrows what to remember from the conversation.

   • Example: $cf-remember the auth flow → filter conversation for auth-related knowledge
   • Example: $cf-remember why we chose Redis → focus on that specific decision
   • Example: $cf-remember gotchas from the migration → extract only gotchas related to migration

2. An explicit topic name — only when the user clearly specifies a filename or category.

   • Example: $cf-remember topic:caching-strategy → use "caching-strategy" as the topic name
   • Example: $cf-remember save as conventions/api-naming → use the explicit path

Default behavior: Always auto-detect the topic name from the filtered content. Only use $ARGUMENTS as the literal topic name when the user explicitly indicates it (e.g., with "topic:", "save as", "call it", "name it").

When $ARGUMENTS acts as a filter:

• Use it to narrow which parts of the conversation to extract knowledge from
• Ignore conversation content that doesn't relate to the user's stated focus
• The topic name should be derived from the filtered content, not from $ARGUMENTS verbatim

Rules

• If no arguments given, scan the entire conversation for key knowledge
• Be concise — bullet points over paragraphs
• Include code snippets only when they clarify the point
• Create directories as needed