PMB vs Voyage AI

Side-by-side comparison of features, pricing, and ratings

Analysis reviewed Live tool data as of 2026-10-18
Cross-checked through our multi-step verification ·
Saved

At a glance

DimensionPMBVoyage AI
PricingFree (open source, Apache 2.0)Contact sales (custom pricing)
Primary use casePersistent memory for AI coding agentsEnterprise RAG with domain-specific embeddings
DeploymentLocal, offline-first, no API keysCloud API
Key integrationsClaude Code, Cursor, Codex, Zed via MCPAny vector DB or LLM (no pre-built lists)
Specialized modelsN/A (context injection via hybrids recall)Voyage-3, finance, legal, code, multimodal
ComplianceNot mentioned (local data = minimal compliance risk)SOC 2, HIPAA

Voyage AI is built for enterprises needing high-accuracy, domain-specific retrieval at scale, while PMB targets developers who want simple, private, local memory for coding agents. Choose Voyage if you run a production RAG system on sensitive data; choose PMB if you're tired of re-explaining context to Claude Code or Cursor.

Try PMBTry Voyage AI
PMB
PMB

Local-first memory layer that gives your AI coding agent a SQLite file it can recall from between sessions.

Visit Website
Voyage AI
Voyage AI

Voyage AI delivers domain-tuned embedding models and rerankers for high-precision RAG retrieval

Visit Website
Pricing
Free
Paid
Plans
$0
Consumption-based pricing (rates not published on page)
Popularity
4 views
7.4k views
Skill Level
Intermediate
Intermediate
API Available
Platforms
CLIDesktop
WebAPI
Categories
🧠 Agent Memory & Runtimes🔌 MCP Servers & Agent Tooling
🗄️ Vector Databases & Retrieval
Features
Local SQLite memory file with vectors in LanceDB stored beside it
Auto-recall: memory injected before the model reasons, via hooks rather than tool calls
MCP server exposing 29 tools over stdio as a child process of your agent
Hybrid recall fusing BM25, dense vectors and an entity graph with Reciprocal-Rank-Fusion
Optional rerank pass on top of fused retrieval results
Sub-millisecond prompt classification on every message
4-16 ms memory reads
~35 ms recall in a real workspace
Async writes: SQLite first, embed and LanceDB insert on a background thread
Lessons carry a surface_id and are scored on agent follow-rate; ignored rules flagged dead
Local web dashboard with interactive entity graph (Map) and git-style Timeline
Entity graph nodes color-coded by type, sized by importance, inspectable on hover
Automatic memory decay, archiving and deduping over time
Optional local Ollama for summarization and graph extraction, fully offline
Multi-project support with one Timeline lane per project
General-purpose embedding models including voyage-3.5 and voyage-3.5 lite
Domain-specific embedding models optimized for finance, legal, and code
Company-specific fine-tuned embedding models on proprietary data
Voyage 4 model series for improved retrieval quality
voyage-multimodal-3.5 embeds images and text in one retrieval pipeline
Low-dimensional embeddings (3x-8x shorter vectors) cut storage and search costs
32K-token long-context support for embedding long documents
rerank-2.5 and rerank-2.5-lite add instruction-following to ranking
voyage-context-3 keeps chunk-level detail with global document context
Batch API for large-scale embedding workloads
4x smaller model with faster inference and superior accuracy
2x cheaper inference with superior accuracy
Plug-and-play with any vectorDB and any LLM
SOC 2 and HIPAA compliance
Deploy on major clouds, in-VPC customer tenants, or on-premise with model licensing
Integrations
Claude Code
Cursor
Codex
Zed
Ollama

What real users say: PMB vs Voyage AI

Not marketing copy and not our opinion — a structured sweep of public discussion (reviews, forums, communities and video comments), showing what people praise and what they complain about for each tool.

PMB

81 mentions across 6 sources · 74% positive (averaged across 6 sources)

Hacker News, Product Hunt, App Store, Bluesky, GitHub, Lemmy

What users praise

  • • Local-first: all data stays on your disk, no cloud or API keys.
  • • Hybrid recall (BM25 + vectors + graph) delivers relevant memory in ~35 ms.
  • • Automatic write and read hooks integrate with Claude Code, Cursor, Codex, Zed.
  • • Honest impact scoring flags dead memories that agents don't follow.

What frustrates them

  • • No cloud or team-synced version, only local storage.
  • • Stale or reversed decisions can still surface without manual pruning.
  • • Automatic memory injection uses context window tokens, reducing solution runway.
  • • Only one maintainer; open issues may take time to resolve.

Researched Jul 4, 2026

Voyage AI

64 mentions across 6 sources · 45% positive — mixed (weighted across 6 sources)

Hacker News, YouTube, App Store, Stack Overflow, GitHub, Lemmy

What users praise

  • • Domain-tuned embedders for finance, legal, and code deliver measurable retrieval precision gains
  • • Harvey case study showed 25% fewer irrelevant documents and 3x lower vectorDB costs
  • • Low-dimensional embeddings (3x-8x shorter vectors) cut storage and search costs materially
  • • Instruction-following rerank-2.5 and rerank-2.5-lite let you steer ranking behavior in plain language

What frustrates them

  • • Terms grant Voyage an irrevocable license to train on API customer data by default
  • • Usage-based pricing at $0.12-$0.18 per million tokens can balloon for high-frequency agents
  • • A community benchmarker found a small Jina model outperformed Voyage for their use case
  • • Circular import bug in voyageai v0.3.5 broke lazy import systems before being fixed

Researched Oct 14, 2026

Who should pick which

  • Enterprise RAG engineer
    Pick: Voyage AI

    Requires domain-specific embedding models (finance, legal) with 32K context, SOC 2/HIPAA compliance, and batch processing for large document corpora.

  • Solo developer building with Claude Code
    Pick: PMB

    Free, local-first persistent memory eliminates repetitive context re-explanation; integrates via MCP with no API keys or latency.

  • Small startup needing vector search on a budget
    Pick: Voyage AI

    Low-dimensional embeddings cut vector storage costs; contact pricing may yield affordable entry, though transparency is lacking.

  • Privacy-conscious developer using Cursor
    Pick: PMB

    All data stays local, no telemetry, no cloud dependency—ideal for proprietary code or sensitive projects.

  • Multimodal retrieval project
    Pick: Voyage AI

    Voyage recently announced voyage-multimodal-3.5, enabling image+text retrieval for RAG pipelines.

Frequently Asked Questions

PMB vs Voyage AI: which should you choose?

Voyage AI is built for enterprises needing high-accuracy, domain-specific retrieval at scale, while PMB targets developers who want simple, private, local memory for coding agents. Choose Voyage if you run a production RAG system on sensitive data; choose PMB if you're tired of re-explaining context to Claude Code or Cursor.

Can I use Voyage AI for free?

No, Voyage AI requires contacting sales for custom pricing; there is no free tier.

Is PMB truly offline?

Yes, PMB runs entirely locally with an SQLite file; no internet connection required after installation.

Which tools does PMB integrate with?

PMB integrates via MCP with Claude Code, Cursor, Codex, and Zed.

Does Voyage AI support multimodal?

Yes, voyage-multimodal-3.5 was recently announced, adding image+text embedding capabilities.

What compliance does Voyage AI offer?

Voyage AI provides SOC 2 and HIPAA compliance for enterprise workloads.

Can I self-host PMB?

Yes, PMB is open source under Apache 2.0; you can inspect and modify the code and run it locally.

Does Voyage AI have a reranker?

Yes, Voyage offers rerank-2.5 and rerank-2.5-lite with instruction following.

What recall method does PMB use?

PMB uses hybrid recall combining BM25, dense vectors, entity graph, and optional rerank, fused with Reciprocal-Rank-Fusion (RRF).

More PMB or Voyage AI comparisons

coreai-model-zoo vs Voyage AI comparison

These aren't competitors, so there's no head-to-head to score. Voyage AI is a paid retrieval layer you call from a RAG pipeline — embeddings, rerankers, and now the Voyage 4 series plus voyage-multimo

AI-Search vs Voyage AI comparison

These two are not competitors and you will never pick one over the other. Voyage AI is a paid retrieval layer — embedding models plus rerankers — for teams whose RAG answers hinge on getting the right

Nubase vs Voyage AI comparison

These aren't substitutes — you would not pick one over the other for the same job. Purchase Voyage AI when retrieval precision is the bottleneck: its domain-tuned models (finance, legal, code), instru

gitlab-duo-provisioning-blueprint vs Voyage AI comparison

Choose Voyage AI if your core need is high-accuracy retrieval on domain-specific data (finance, legal) with long-context support and low storage costs. Choose gitlab-duo-provisioning-blueprint if you

sandboxd vs Voyage AI comparison

These are not competitors, so don't treat it as a head-to-head. Buy Voyage AI if your problem is retrieval precision: you're paying per token for embedders and rerankers that carry finance, legal, and

agentteam-email vs Voyage AI comparison

These aren't competitors — don't make a buyer choose between them. Voyage AI is a retrieval-layer purchase: you're buying embedding and reranking accuracy for RAG, with domain-tuned models for finance

Explore each tool further

PMB
View PMB reviewPMB alternatives
Voyage AI
View Voyage AI reviewVoyage AI alternatives

Browse these categories

Best AI Agent Memory & Runtimes toolsBest AI Vector Databases & Retrieval toolsBest AI MCP Servers & Agent Tooling tools
Still deciding? Get the weekly AI tools brief

One email a week — new tools, honest comparisons, no spam.

Last reviewed: July 2, 2026