help

/help — Memex Guide & Status

What is memex?

Memex injects relevant knowledge into your 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, rules, and memories. The top matches are injected as additional context. Match counts are tracked via telemetry (~/.claude/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
/sleep	Knowledge lifecycle: migrate CLAUDE.md/MEMORY.md/rules into skills, promote/demote entries based on match telemetry
/deep-sleep	Analyze past session transcripts to extract recurring patterns, classify into appropriate types
/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
/takeover	Read a handoff document and resume the work
/help	This guide

Creating content

Skills

Create ~/.claude/skills/<name>/SKILL.md (global) or <project>/.claude/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

Create ~/.claude/rules/<name>.md (global) or <project>/.claude/rules/<name>.md (project):

---
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.

Rules without frontmatter work too — filename becomes name, first line becomes description.

Memories

Memories live in ~/.claude/projects/<encoded-cwd>/memory/*.md (where <encoded-cwd> is the cwd with / → - and . → -).

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 /sleep to convert accumulated MEMORY.md entries into searchable skills, and /reflect to extract learnings from the current conversation.

Check status

Run these to inspect memex's current state:

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

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

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

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

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

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

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

Configuration

Create ~/.claude/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": {}
  },
  "sleepSchedule": {
    "enabled": false,
    "dailyAt": "03:00",
    "projects": []
  },
  "hooks": {
    "UserPromptSubmit": {
      "enabled": true,
      "topK": 3,
      "threshold": 0.5,
      "maxInjectedChars": 8000,
      "types": ["skill", "memory", "workflow", "session-learning", "rule"]
    },
    "PreToolUse": {
      "enabled": false,
      "topK": 2,
      "threshold": 0.6,
      "maxInjectedChars": 4000,
      "types": ["tool-guidance", "skill"]
    },
    "Stop": {
      "enabled": false,
      "extractLearnings": true,
      "behavioralRules": true
    }
  }
}

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)
• sync — enable cross-machine sync via a private git repo (see README)

Troubleshooting

Symptom	Check
Nothing is being injected	Are there skills/rules/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
Queries feel stale or ineffective	Run /sleep — it evolves weak queries based on queryHits telemetry
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 ~/.claude/cache/memex-cache.json to force rebuild
Hook not running at all	Check ~/.claude/settings.json for the hook registration

$ARGUMENTS