VEX — Vector Exchange

Cross-standard vector DB migration, memory portability, and sovereign backup for AI agents.

Move memories between any vector store. Import Claude and ChatGPT conversation history with LLM fact extraction. Back up encrypted memory to GitHub, Codeberg, or self-hosted Gitea. Convert to any LLM provider format.

npm install -g @vektormemory/vex

Commands

| Command | Description | |---------|-------------| | vex sync | Sovereign hybrid backup — encrypt and push memory to GitHub, Codeberg, or Gitea | | vex migrate | Migrate directly between any two vector stores | | vex export | Export memory to .vmig.jsonl portable format | | vex import | Import .vmig.jsonl into any store | | vex convert | Convert to OpenAI, Anthropic, Groq, or plain text format | | vex sign | Sign export with BLAKE3 + Ed25519 | | vex verify | Verify export signature | | vex inspect | Show stats, namespaces, dimensions | | vex validate | Lint all records in a .vmig.jsonl file |

Connectors

| Store | Import | Export | Notes | |-------|--------|--------|-------| | VEKTOR Slipstream | ✓ | ✓ | Local SQLite, full MAGMA graph | | Pinecone | ✓ | ✓ | Serverless and pod-based | | Qdrant | ✓ | ✓ | Local and Cloud | | ChromaDB | ✓ | ✓ | Local and hosted | | Weaviate | ✓ | ✓ | Local and Cloud | | pgvector | ✓ | ✓ | PostgreSQL, Supabase compatible | | Redis | ✓ | ✓ | RediSearch VSS, plain hash fallback | | Milvus / Zilliz | ✓ | ✓ | Local Milvus and Zilliz Cloud | | Neo4j | ✓ | ✓ | Local and Aura Cloud | | Claude export | ✓ | — | conversations.json from Claude | | ChatGPT export | ✓ | — | conversations.json from ChatGPT | | .vmig.jsonl | ✓ | ✓ | Portable interchange format |

Sovereign Backup — vex sync

Back up VEKTOR memory to any Git host. Encrypted client-side with AES-256-GCM. Cloud stores ciphertext only — no plaintext memory ever leaves your machine.

Supported providers:

| Provider | Cost | Notes | |----------|------|-------| | GitHub | Free private repos | Familiar, reliable | | Codeberg | Free, nonprofit | GDPR-compliant, no tracking — recommended | | Gitea | Self-hosted | Full sovereignty, single binary, runs on any VPS | | GitLab | Free tier | gitlab.com or self-hosted |

# Initialize with Codeberg (recommended — free, GDPR, nonprofit)
vex sync init --provider codeberg \
  --token cb_xxx \
  --owner alice \
  --repo vektor-backup \
  --db ~/.vektor/slipstream-memory.db

# Initialize with GitHub
vex sync init --provider github \
  --token ghp_xxx \
  --owner alice \
  --repo vektor-backup \
  --db ~/.vektor/slipstream-memory.db

# Initialize with self-hosted Gitea (full sovereignty)
vex sync init --provider gitea \
  --gitea-url https://git.example.com \
  --token xxx \
  --owner alice \
  --repo vektor-backup \
  --db ~/.vektor/slipstream-memory.db

# Push all memories (encrypted)
vex sync push

# Push only high-importance memories (lean backup)
vex sync push --min-importance 3

# Check sync status
vex sync status

# Compare local vs remote
vex sync diff

# Restore to new machine
vex sync pull --db ~/new-machine-memory.db

# Dry run restore (preview without writing)
vex sync pull --db ~/memory.db --dry-run

How encryption works:

• Key = HKDF-SHA256(machine-id + token hash) — derived locally, never transmitted
• Payload = AES-256-GCM encrypted export — blob is opaque to the Git host
• Manifest = plaintext metadata only (memory count, timestamp) — no content
• To restore on a new machine: provide your original token + re-run vex sync init

Quick Start

Import Claude conversation history with fact extraction

# Smart mode: extracts facts using LLM, scores by importance, deduplicates
vex migrate --from claude-export --to vektor \
  --file conversations.json \
  --db memory.db \
  --mode smart \
  --groq-key $GROQ_KEY \
  --namespace my-history

# With OpenAI (recommended for large exports)
vex migrate --from claude-export --to vektor \
  --file conversations.json \
  --db memory.db \
  --mode smart \
  --openai-key $OPENAI_KEY \
  --extract-model openai:gpt-4o-mini \
  --concurrency 5

Migrate between vector stores

# VEKTOR → Pinecone
vex migrate --from vektor --to pinecone \
  --db memory.db \
  --api-key $PINECONE_KEY \
  --index my-index \
  --host $PINECONE_HOST

# VEKTOR → Redis (with RediSearch VSS)
vex migrate --from vektor --to redis \
  --db memory.db \
  --redis-url redis://localhost:6379 \
  --index vex-memory

# VEKTOR → Neo4j (builds full knowledge graph)
vex migrate --from vektor --to neo4j \
  --db memory.db \
  --neo4j-url bolt://localhost:7687 \
  --neo4j-user neo4j \
  --neo4j-password secret

# VEKTOR → Milvus
vex migrate --from vektor --to milvus \
  --db memory.db \
  --milvus-url localhost:19530 \
  --collection vex_memory

# VEKTOR → PostgreSQL + pgvector
vex migrate --from vektor --to pgvector \
  --db memory.db \
  --url postgres://user:pass@localhost:5432/mydb

Export and convert

# Export to portable vmig format
vex export --from vektor --db memory.db --output memories.vmig.jsonl

# Convert for OpenAI fine-tuning
vex convert --from memories.vmig.jsonl --adapter openai-finetune --output finetune.jsonl

# Convert for Anthropic Messages API
vex convert --from memories.vmig.jsonl --adapter anthropic-finetune --output anthropic.jsonl

# Convert for Groq / Mistral / Together / Perplexity
vex convert --from memories.vmig.jsonl --adapter generic-chat --output chat.jsonl

# Plain text transcript
vex convert --from memories.vmig.jsonl --adapter plain-text --output transcript.txt

Sign and verify exports

# Sign for tamper-evident transfer (BLAKE3 + Ed25519)
vex sign memories.vmig.jsonl

# Verify before importing
vex verify memories.vmig.jsonl && vex import --from memories.vmig.jsonl --to qdrant --collection mem

Extraction Pipeline

When using --mode extract or --mode smart, VEX runs a 7-step pipeline:

1. Chunk — split conversations by mode (turn, conversation, exchange)
2. Extract — LLM fact extraction with importance scoring, tags, and potential questions
3. Score — filter by importance threshold (default 0.5)
4. Dedup — remove near-duplicates (default threshold 0.72)
5. Embed — generate vectors via OpenAI, Ollama, or custom endpoint
6. Graph — build temporal, tag, and causal edges between facts
7. Store — write to target vector store

Provider cascade

# Auto-detect from config
vex migrate ... --provider auto

# Explicit cascade: Groq first, fall to Ollama
vex migrate ... --provider groq,ollama --groq-key $KEY --ollama-url http://localhost:11434

# Key rotation: 3 Groq keys triples effective TPM budget
vex migrate ... --provider groq --groq-key $KEY1,$KEY2,$KEY3

# Override model per provider
vex migrate ... --extract-model groq:llama-3.3-70b-versatile,ollama:mistral

Extraction Flags

| Flag | Default | Description | |------|---------|-------------| | --mode | raw | raw, extract, or smart | | --provider | auto | groq, ollama, openai, anthropic, mistral, together | | --groq-key | — | Groq API key(s), comma-separated for rotation | | --openai-key | — | OpenAI API key | | --anthropic-key | — | Anthropic API key | | --ollama-url | localhost:11434 | Ollama endpoint | | --extract-model | provider default | Override model, e.g. groq:llama-3.3-70b-versatile | | --min-importance | 0.5 | Filter threshold | | --concurrency | 3 | Parallel LLM calls (use 1 for free Groq tier) | | --dry-run | — | Preview facts without storing | | --namespace | — | Tag all stored records with namespace |

The vmig Format

.vmig.jsonl is VEX's portable interchange format. Each line is a JSON record:

{
  "id": "uuid",
  "text": "The fact or memory content",
  "vector": [0.1, 0.2, ...],
  "namespace": "my-history",
  "source": "vektor",
  "created_at": 1780525199,
  "metadata": {
    "importance": 0.85,
    "tags": ["decision", "architecture"],
    "potential": [
      "What database did Mini choose?",
      "Why was SQLite selected over Postgres?",
      "What storage approach does VEKTOR use?"
    ]
  }
}

Files can be inspected, validated, and signed without a vector store:

vex inspect memories.vmig.jsonl   # stats, namespaces, dimensions
vex validate memories.vmig.jsonl  # lint all records
vex sign memories.vmig.jsonl      # BLAKE3 + Ed25519 signature
vex verify memories.vmig.jsonl    # exit 0=valid, 1=tampered

Install Optional Connector Dependencies

# Redis
npm install redis

# Milvus / Zilliz Cloud
npm install @zilliz/milvus2-sdk-node

# Neo4j / Aura
npm install neo4j-driver

# PostgreSQL / pgvector / Supabase
npm install pg

# LangChain integration
npm install @langchain/core

Related

• VEKTOR Slipstream — local-first persistent memory SDK for AI agents

License

Apache 2.0 — vektormemory.com