help

name	help
description	Show how to use memex: what it does, how to create skills/rules/memories, available commands, current status, and troubleshooting.
queries	["how does memex work","how to use memex","what skills are available","how to create a skill","how to create a rule","memex help","memex is not working","no skills are being injected"]

/help — Memex Guide & Status

What is memex?

Memex injects relevant knowledge into your Codex session based on what you're working on. Instead of loading everything into context at once, it uses semantic similarity to surface only what's needed right now.

How it works

Each prompt you type is embedded locally (ONNX, no API calls) and compared against your indexed skills and memories. The top matches are injected as additional context via the UserPromptSubmit hook. Match counts are tracked via telemetry (~/.codex/cache/memex-telemetry.json).

Entry types and disclosure model:

Type	First match	Subsequent matches	Matched on
rule	Full content	one-liner reminder only	UserPromptSubmit
memory	Full content	Full content	UserPromptSubmit
skill	Description teaser	Full content (via Read)	UserPromptSubmit
workflow	Description teaser	Full content (via Read)	UserPromptSubmit
tool-guidance	Full content	Full content	PreToolUse
stop-rule	Behavioral rules	—	Stop
session-learning	Full content	Full content	UserPromptSubmit

Bundled skills

Command	Description
/deep-sleep	Drain the Stop-hook learnings queue and analyze rollout JSONL transcripts to extract recurring patterns
/reflect	Extract learnings from the current conversation and save as memories, rules, or skills
/doctor	Diagnose installation and configuration issues
/handoff	Write a continuation plan to disk so a fresh session can pick up where this one left off
/help	This guide

Creating content

Skills

Create ~/.codex/skills/<name>/SKILL.md (global) or <project>/.agents/skills/<name>/SKILL.md (project):

---
name: my-skill
description: "What this skill does"
queries:
  - "when would someone need this"
  - "another example query"
boost: 0.05  # optional: nudge score for entries near the threshold
---
The skill content.

Rules

Rules are skills with type: rule frontmatter:

---
name: my-rule
description: "What this rule enforces"
type: rule
one-liner: "Short reminder for subsequent matches"
queries:
  - "when this rule applies"
---
Full rule content shown on first match.

Memories

Memories live in ~/.codex/memex/projects/<encoded-cwd>/memory/*.md (where <encoded-cwd> is the cwd with / → -, . → -, _ → -).

Use ## headings for each entry and optionally add Triggers: lines for better semantic matching:

## Prefer pnpm
Triggers: install dependencies, npm install, which package manager
Use `pnpm` instead of `npm` for all operations.

Use /reflect to extract learnings from the current conversation, and /deep-sleep to process queued sessions from past Codex rollouts.

Check status

Run these to inspect memex's current state:

# Config (or defaults if no file)
cat ~/.codex/memex.json 2>/dev/null || echo "Using defaults"

# What's indexed — global
ls ~/.codex/skills/*/SKILL.md 2>/dev/null

# What's indexed — this project
ls .agents/skills/*/SKILL.md 2>/dev/null

# Memories for this project
ls ~/.codex/memex/projects/*/memory/*.md 2>/dev/null

# Pending learnings queue (Stop hook capture)
cat ~/.codex/cache/memex-learnings-queue.json 2>/dev/null

# Cache status
ls -la ~/.codex/cache/memex-cache.json 2>/dev/null

# Match telemetry
cat ~/.codex/cache/memex-telemetry.json 2>/dev/null

# Model cache
ls ~/.codex/cache/models/ 2>/dev/null

# Plugin install
codex plugin list 2>/dev/null | grep memex

Configuration

Create ~/.codex/memex.json to customize. All fields optional — defaults shown:

{
  "enabled": true,
  "embeddingModel": "Xenova/all-MiniLM-L6-v2",
  "cacheTimeMs": 300000,
  "skillDirs": [],
  "sync": {
    "enabled": false,
    "repo": "",
    "autoPull": true,
    "autoCommitPush": true,
    "projectMappings": {}
  },
  "hooks": {
    "UserPromptSubmit": {
      "enabled": true,
      "topK" 
       
       
           
    
     
       
       
       
       
        
    
     
       
       
       
    
  

Key tuning knobs:

• threshold — lower to match more broadly (default 0.5), raise for precision
• topK — how many matches to inject per prompt (default 3)
• maxInjectedChars — character budget for injected content (default 8000)
• hooks.Stop.extractLearnings — enqueue sessions to memex-learnings-queue.json for /deep-sleep
• sync — enable cross-machine sync via a private git repo (see README)

Troubleshooting

Symptom	Check
Nothing is being injected	Are there skills/memories in the scan paths? (see "Check status" above)
Wrong things are injected	Adjust threshold higher, or improve queries in your SKILL.md frontmatter
Skill barely misses threshold	Add boost: 0.05 to frontmatter to nudge its score up
Everything is injected	threshold may be too low, raise it (e.g. 0.6)
Slow first run	Model download (~23MB), needs internet access
Stale results after editing skills	Delete ~/.codex/cache/memex-cache.json to force rebuild
Hook not running at all	Run /doctor — verify plugin installed and hooks trusted via /hooks
Learnings not extracted	Enable hooks.Stop.enabled and extractLearnings in ~/.codex/memex.json

$ARGUMENTS