reflect-ingest

name	reflect:ingest
description	The global knowledge indexer. Harvests ALL memory sources across all tools (Claude, Codex, Copilot, Gemini) and all project types into the unified GraphRAG + QMD knowledge base. Archives originals, generates entity sidecars, and dual-indexes for future retrieval. This is THE command that makes the knowledge base comprehensive.
version	3.0.0
user-invocable	true
triggers	["reflect:ingest","ingest memories","index learnings","harvest memories","build knowledge base"]
allowed-tools	["Read","Write","Edit","Grep","Glob","Bash"]

Reflect: Ingest — Global Knowledge Indexer

The single command that makes the knowledge base comprehensive. Sweeps ALL memory sources across all tools and projects, archives originals, generates entity sidecars, and dual-indexes into GraphRAG + QMD.

Philosophy: Every piece of knowledge from every tool should be searchable in one place. If it exists in a memory file somewhere, it should be in the global index.

When to Use

• Periodically (weekly recommended) to keep the knowledge base current
• Before deleting projects or cleaning up ~/.claude/projects/
• After heavy multi-project work across different tools
• When reflect:status shows low sidecar coverage or stale sources
• After running reflect:consolidate (to index what it produced)

What Gets Indexed

reflect:ingest sweeps ALL of these sources:

#	Source	Path	Provider
1	Claude auto-memories (individual files)	~/.claude/projects/*/memory/*.md	claude.py
2	Codex memories	~/.codex/memories/*.md	codex.py
3	Codex global AGENTS.md	~/.codex/AGENTS.md	codex.py
4	Copilot AGENTS.md	~/.copilot/AGENTS.md	copilot.py
5	Gemini global config	~/.gemini/GEMINI.md	gemini.py
6	Gemini project configs	<project>/GEMINI.md	gemini.py
7	Project AGENTS.md files	<repo>/AGENTS.md	codex.py / copilot.py
8	Consolidated project memory	<repo>/.agents/MEMORY.md	Glob scan
9	Unindexed knowledge notes	docs/solutions/**/*.md without sidecars	Glob scan
10	Episode files	~/.reflect/episodes/ep-*.md	Glob scan
11	Skill-embedded learnings	~/.claude/skills/*/references/*.md	Glob scan

Output Destinations

Everything flows to:

~/.learnings/                (KB content + index; managed by the `reflect` CLI)
├── documents/
│   ├── *.md / *.entities.yaml  Learning notes + entity sidecars (indexed
│   │                            into GraphRAG + QMD)
│   ├── memories/               Archived originals (by project)
│   │   ├── shotclubhouse/
│   │   ├── ai-coder-rules/
│   │   └── ...
│   └── episodes/               Session episode notes
├── nano_graphrag_cache/        GraphRAG index (nodes, edges, communities)
├── shards/                     R15: per-project shard KBs (each its own
│   ├── <project-a>/            documents/ + nano_graphrag_cache/). Recall
│   │   ├── documents/          defaults to the CURRENT project's shard for
│   │   ├── nano_graphrag_cache/ faster, cross-project-noise-free injection;
│   │   └── branches/           A6: per-branch / per-worktree sub-shards.
│   │       ├── feat__auth/     Worktrees of one repo (agents-in-a-box runs
│   │       │   ├── documents/  /.agents-in-a-box/worktrees/...) get SEPARATE
│   │       │   └── nano_graphrag_cache/ sub-shards so feat/auth and
│   │       └── feat__payment/  feat/payment don't pollute each other's recall.
│   └── <project-b>/            Recall defaults to the CURRENT branch's sub-
│                                shard; trunk (main/master) uses the project
│                                shard. `reflect:recall --all-branches` pools
│                                every branch of the project; `--global` pools
│                                every project at the top-level KB.
└── .memory-ingest-log.yaml     Tracks what's been ingested (prevents reprocessing)

# The `reflect` CLI itself is installed separately via:
#   uv tool install --upgrade --torch-backend cpu 'git+https://github.com/stevengonsalvez/ainb-reflect-memory.git[graph]'
# and lands on $PATH at ~/.local/bin/reflect. The legacy ~/.learnings/cli/
# install path is deprecated; the `reflect` binary fully supersedes it.

~/.cache/qmd/
├── index.sqlite            QMD search index
└── (collections: learnings, obsidian, blog, writing)

Pipeline

Step 1: Discover ALL sources

# Use multi-tool provider discovery
python3 {{HOME_TOOL_DIR}}/skills/reflect/scripts/memory_discovery.py discover --json

# Also scan non-provider sources
# Episode files:
find ~/.reflect/episodes -name "ep-*.md" 2>/dev/null

# Unindexed knowledge notes (missing sidecars):
find docs/solutions -name "*.md" -not -name "*.entities.yaml" 2>/dev/null | while read f; do
  [ ! -f "${f%.md}.entities.yaml" ] && echo "$f"
done

# .agents/MEMORY.md in current repo:
[ -f .agents/MEMORY.md ] && echo ".agents/MEMORY.md"

Step 2: Read and Classify Each Memory File

For each discovered file:

1. Read the file fully
2. Extract frontmatter fields: name, description, type
3. Classify by memory type → learning route:

Memory Type	Learning Type	Scope	Confidence
feedback	correction / pattern	universal (if general) or project	HIGH (user stated)
user	preference	universal	HIGH
project	decision / pattern	project	MEDIUM
reference	reference	project or domain	MEDIUM
Episode	session summary	project	MEDIUM
Knowledge note	(already structured)	from frontmatter	from frontmatter
AGENTS.md section	preference / process	universal	HIGH

4. Extract project name from directory path:

   -Users-stevengonsalvez-d-git-shotclubhouse → shotclubhouse
   

Step 3: Dedup Against Existing Knowledge Base

Check each entry against what's already indexed:

# Check QMD first (fastest)
qmd query --collection learnings --json "{key_insight}" 2>/dev/null

# Check GraphRAG via the reflect CLI (from reflect-kb — installed via
# `uv tool install reflect-kb`). Older versions of this skill referenced
# a legacy $HOME/.learnings/cli/learnings wrapper — that path is gone
# after the global-learnings skill purge; use `reflect` instead.
reflect search "{key terms}" 2>/dev/null

# Check ingest log (prevents reprocessing)
grep -q "{content_hash}" "$HOME/.learnings/.memory-ingest-log.yaml" 2>/dev/null

Dedup rules:

• Same content_hash in ingest log → SKIP (already processed)
• High-similarity match in QMD (>85%) → SKIP (duplicate)
• Same topic but different detail level → KEEP (more detail supersedes)
• Same error + same fix → SKIP
• Different angle on same topic → KEEP both, note relationship

Step 4: Convert to Knowledge Learnings

For each non-duplicate memory, generate:

A. Learning note (~/.learnings/documents/{id}.md — flat top-level; the CLI chooses this destination):

Use the template at ${CLAUDE_PLUGIN_ROOT}/plugin/assets/learning_template.md for the note structure. Fields preserved: id, scope, confidence, learning_type, source_episodes, superseded_by (for revisions), provenance (source_tool/path/hash), plus Problem / Solution / Anti-Pattern / Context sections.

---
title: "{descriptive title}"
category: "{auto-detected category}"
type: LEARNING
scope: "{universal|project|domain}"
confidence: {0.0-1.0}
key_insight: "{the one-line takeaway}"
tags: ["{project}", "{tech}", "{topic}"]
provenance:
  source_tool: "{claude|codex|copilot|gemini}"
  source_path: "{original file path}"
  content_hash: "{sha256 prefix}"
  ingested_at: "{ISO timestamp}"
---

## Problem
{context from the memory — what situation triggered this}

## Solution
{the actionable guidance from the memory body}

## Context
{project name, source path, related files}

B. Entity sidecar (~/.learnings/documents/{id}.entities.yaml — sibling of the learning note):

Required schema — every field below is mandatory except strength (defaults to 5). The reflect add --entities CLI enforces this via entity_store.Entity and will raise KeyError if any required key is missing.

document_id: lrn-{slug}-{hash6}
extracted_at: "{ISO-8601 timestamp}"
entities:
  - name: "{canonical lowercase name}"
    type: technology | error | pattern | function | concept | tool
    description: "{brief one-line description}"      # REQUIRED
relationships:
  - source: "{entity A name}"                        # NOT 'from'
    target: "{entity B name}"                        # NOT 'to'
    type: caused_by | causes | enables | prevents | contradicts | supersedes | part_of | uses | solves | requires | relates_to | implements | configures | triggers
    description: "{how they relate}"                 # REQUIRED
    strength: 1-10                                   # optional, default 5

Entity type hints by memory kind:

• Tools, frameworks, libraries → technology
• Patterns, anti-patterns → pattern
• Errors, gotchas → error
• Functions, APIs → function
• Concepts → concept

Relationship hints by memory kind:

• feedback: solves or relates_to
• project: requires or relates_to
• reference: relates_to to external resources
• when the direction of effect is known, prefer a typed causal link (S2) — caused_by / causes / enables / prevents / contradicts / supersedes / part_of / uses — over flat relates_to

Validate before writing:

python3 {{HOME_TOOL_DIR}}/skills/reflect/scripts/validate_sidecar.py "$SIDECAR_PATH"

See ${CLAUDE_PLUGIN_ROOT}/plugin/references/knowledge_format.md for the full entity/relationship type reference.

Step 5: Present Summary for Approval

## Ingest Plan

### Sources Scanned
| Provider | Files Found | New | Duplicate | Stale |
|----------|------------|-----|-----------|-------|
| Claude   | 45         | 12  | 30        | 3     |
| Codex    | 8          | 3   | 5         | 0     |
| Copilot  | 2          | 0   | 2         | 0     |
| Gemini   | 4          | 2   | 2         | 0     |
| Episodes | 9          | 2   | 7         | 0     |
| docs/    | 3          | 3   | 0         | 0     |

### New Learnings to Index
| # | Source | Project | Title | Type | Category |
|---|--------|---------|-------|------|----------|
| 1 | claude | shotclubhouse | "RLS bypass with SECURITY DEFINER" | pattern | database |
| 2 | codex  | ai-coder-rules | "Codex memories dir structure" | reference | tooling |
| 3 | episode | — | "Sprint bug blitz session" | session | debugging |
...

**New learnings**: 22
**Skipped (duplicates)**: 46
**Sidecars to generate**: 3 (for existing unindexed notes)

Proceed? [Y/n/select by number]

NEVER auto-index without user approval.

Step 6: Archive Originals and Index

For each approved entry:

A. Archive original (preserves raw content in git):

LEARNINGS_HOME="${LEARNINGS_HOME:-$HOME/.learnings}"
PROJECT_NAME="$(decode_project "$PROJECT_SLUG")"
ARCHIVE_DIR="$LEARNINGS_HOME/documents/memories/$PROJECT_NAME"
mkdir -p "$ARCHIVE_DIR"

# Copy with archive metadata
{
  echo "<!-- archived: $(date -Iseconds) -->"
  echo "<!-- source: $ORIGINAL_PATH -->"
  echo ""
  cat "$ORIGINAL_PATH"
} > "$ARCHIVE_DIR/$(basename "$ORIGINAL_PATH")"

B. Write learning note + entity sidecar

C. Index into GraphRAG + QMD:

# Use the canonical reflect CLI (from reflect-kb, installed via
# `uv tool install reflect-kb`). --force is REQUIRED here because this
# step runs under a non-interactive subprocess; without it, an existing
# doc with a colliding id would silently abort the add.
reflect add "$NOTE_PATH" --entities "$SIDECAR_PATH" --force

D. Update QMD:

qmd update --collection learnings 2>/dev/null
qmd embed 2>/dev/null

Step 7: Mark as Ingested

Append to tracking file so already-ingested memories aren't reprocessed:

INGEST_LOG="$HOME/.learnings/.memory-ingest-log.yaml"
# Append entry:
# - file: {original path}
#   content_hash: {sha256}
#   ingested_at: {ISO timestamp}
#   learning_id: {generated id}
#   source_tool: {provider name}

Step 8: Git Commit Archive + Learnings

cd "$LEARNINGS_HOME"
git add documents/memories/ documents/learnings/ .memory-ingest-log.yaml
git commit -m "reflect:ingest — $COUNT learnings from $PROVIDER_COUNT providers

Archived: $ARCHIVED original memory files
Indexed: $INDEXED into GraphRAG + QMD
Sources: $SOURCES"

Step 9: Update Metrics

python3 {{HOME_TOOL_DIR}}/skills/reflect/scripts/metrics_updater.py \
    --knowledge-notes $COUNT --sidecars $COUNT

Also rebuild the skills index (R20) so retrieval can match queries against installed skills without a filesystem scan:

python3 {{HOME_TOOL_DIR}}/skills/reflect/scripts/skill_index.py rebuild

Step 10: Report

## Ingest Complete

- **Sources scanned**: 4 providers + episodes + docs
- **New learnings indexed**: 22 (with entity sidecars)
- **Duplicates skipped**: 46
- **Sidecars backfilled**: 3
- **Archives written**: 22 → ~/.learnings/documents/memories/
- **GraphRAG**: {node_count} nodes, {edge_count} edges
- **QMD**: {file_count} files, {vector_count} vectors

Difference from reflect:consolidate

	reflect:consolidate	reflect:ingest
Scope	Current project only	ALL projects, ALL tools
Purpose	Tidy worktree orphans → .agents/MEMORY.md	Build comprehensive global knowledge base
Output	.agents/MEMORY.md (project, git-tracked)	~/.learnings/documents/ (global GraphRAG + QMD)
Cleanup	Deletes orphaned worktree dirs	Marks as ingested (preserves originals as archives)
When	After deleting worktrees	Periodically, or before project cleanup

Typical workflow: Run reflect:consolidate first (tidies project), then reflect:ingest (indexes everything globally).

Safety

• NEVER auto-index without user approval
• NEVER delete original memory files (archive them instead)
• Always check ingest log to prevent reprocessing
• Show full summary before any writes
• Preserve provenance metadata on every learning