# RAG Tracking

> External persistent memory for document ingestion and issues, designed for agents without built-in memory (Claude Code, Codex, OpenCode). Use when ingesting documents, tracking parsing problems, or recalling collection state. Works with ctx for cross-session persistence. NOT needed for Letta Code which has native memory.

- Skill: `etalab-ia/rag-tracking` (Agent Skill, multi-file: 4 files)
- Install (CLI): `npx skillmds@latest add etalab-ia/rag-tracking`
- Raw SKILL.md: https://api.skillmd.com/api/skills/etalab-ia/rag-tracking/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- License: MIT
- Author: etalab-ia (https://skillmd.com/u/etalab-ia)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/etalab-ia/rag-tracking

---


# rag-tracking Skill

External persistent memory for document ingestion and issues, designed for agents without built-in memory. This skill uses **ctx** under the hood for cross-session persistence.

## When to Use This Skill

**Use this skill when:**
- Working with agents that lack persistent memory (Claude Code, Codex, OpenCode, etc.)
- Sharing ingestion state across multiple agents or team members
- Running in CI/CD environments where agent memory doesn't persist

**NOT needed for:**
- Letta Code — has built-in persistent memory (system, progressive, recall)
- Any agent platform with native long-term memory support

For Letta Code users, simply store collection state in your agent's memory files instead.

## Memory Files

### ctx Default Files (created by `ctx init`)

| File | Purpose | When to Use |
|------|---------|-------------|
| `TASKS.md` | Active work | Track what we're doing now |
| `DECISIONS.md` | Architecture decisions | Record why we chose X |
| `LEARNINGS.md` | Patterns learned | Store discovered patterns |
| `CONVENTIONS.md` | Coding conventions | Define how we work |

### Dragster-Specific Files (our templates)

| File | Purpose | When to Use |
|------|---------|-------------|
| `COLLECTIONS.md` | Collection state + ingestion history | Before/after ingesting documents |
| `ISSUES.md` | Parsing problems + workarounds | When encountering errors |

## Location

These files live in `.context/` (ctx's directory) for cross-session persistence:

```
.context/
├── COLLECTIONS.md     # What we've indexed
├── ISSUES.md          # What's broken
├── TASKS.md           # (ctx default) Active work
├── DECISIONS.md       # (ctx default) Architecture decisions
├── LEARNINGS.md       # (ctx default) Patterns learned
└── CONVENTIONS.md     # (ctx default) Conventions
```

## Templates vs Runtime Files

The files in this skill directory (`skills/memory/`) are **templates** for reference.

**Runtime files** live in `.context/` (managed by ctx):

```
.context/
├── COLLECTIONS.md     # Created by ctx init or copied from template
├── ISSUES.md          # Created when first issue is recorded
└── ...
```

### Initial Setup

1. Install ctx: `brew install worktrunk` (or see ctx.ist)
2. Initialize ctx in your project: `ctx init`
   - This creates `.context/` with TASKS.md, DECISIONS.md, LEARNINGS.md, CONVENTIONS.md
3. Copy dragster templates to `.context/`:
   ```bash
   cp skills/memory/COLLECTIONS.md .context/
   cp skills/memory/ISSUES.md .context/
   ```
4. Copy ctx configuration (optional, enables hooks):
   ```bash
   cp skills/memory/ctxrc.template .ctxrc
   ```
5. Add dragster-specific patterns to LEARNINGS.md:
   ```bash
   ctx add learning "French government PDFs require --ocr-language fra"
   ctx add learning "Scanned documents benefit from --dpi 300"
   ```

---

## Enable/Disable

Set `enabled: false` in the frontmatter to disable this skill.

When disabled, the skill won't be loaded by Letta Code and hooks won't fire.

To re-enable: change back to `enabled: true`.

---

## Commands

### Record Ingestion

After successfully ingesting documents, add an entry to COLLECTIONS.md and update the Collections table.

### Record Issue

When encountering a parsing problem, add an entry to ISSUES.md with the error and any workaround found.

### Promote to Learning

When a pattern is discovered, promote to LEARNINGS.md via:
```bash
ctx add learning "<pattern-description>"
```

---

## Workflow

### Starting a Session

1. Read `.context/COLLECTIONS.md` to understand current state
2. Check `.context/ISSUES.md` for any open problems
3. Proceed with document processing

### After Ingestion

1. Update `COLLECTIONS.md` with ingestion details
2. If issues occurred, add entries to `ISSUES.md`
3. If patterns discovered, promote to `LEARNINGS.md`

### Session Wrap-up

1. Review open issues in `ISSUES.md`
2. Mark resolved issues
3. Ensure collections table is accurate

---

## Integration with ctx

This skill integrates with ctx's memory system:

- `ctx agent` includes COLLECTIONS.md and ISSUES.md in context packets
- `ctx add learning` promotes patterns to persistent memory
- `ctx status` shows memory file health

---

## Hooks

Hooks provide automatic reminders at key points in the agent lifecycle.

### Available Hooks

| Hook | When | Reminder |
|------|------|----------|
| `session_start` | Session begins | Load memory context |
| `post_ingestion` | After `lit parse` or `qmd` | Record ingestion |
| `pre_compact` | Before context compaction | Save state |

### Configuration

Hooks are configured in `.ctxrc` (copy from `skills/memory/ctxrc.template`):

```yaml
# .ctxrc
[hooks.session_start]
reminder = "📦 Memory: Read .context/COLLECTIONS.md and .context/ISSUES.md"

[hooks.post_tool_use]
match = "lit parse|qmd collection"
reminder = "Record this ingestion in .context/COLLECTIONS.md"

[freshness_files]
  - path: .context/COLLECTIONS.md
    desc: Document collection state
  - path: .context/ISSUES.md
    desc: Parsing problems and workarounds
```

### Disabling Hooks

Set `enabled: false` in this skill's frontmatter to disable all hooks.