Vectorize
Vectorize offers open-source AI agent memory (Hindsight) and RAG pipelines for turning unstructured data into vector search indexes.
Last verified:
What is Vectorize?
Hindsight is an open-source agent memory system from Vectorize that enables AI agents to learn from experience, recall relevant information, and improve over time. It treats memory as a first-class substrate for reasoning rather than a thin retrieval layer, using three core operations: Retain (convert interactions into structured, time-aware memory), Recall (retrieve the most relevant memories within a token budget), and Reflect (reason over memory to answer questions and update beliefs).
Key features include four parallel retrieval strategies (dense vector search for semantic similarity, sparse vector search for keyword matching, graph traversal for entity relationships, and temporal search for time-aware retrieval), token budgets instead of top-K for predictable context usage, conflict detection that preserves both states of changing facts with timestamps, entity resolution for batch disambiguation, and structured memory organized into distinct networks for facts, experiences, observations, and opinions. The system achieves 94.6% on LongMemEval, the peer-reviewed benchmark for agent memory systems.
Hindsight is designed for developers and teams building AI agents, copilots, and automation that need to operate over long time horizons, maintain consistency across sessions, track evolving user preferences, and learn from mistakes. It works with any LLM (model-agnostic), integrates with MCP-capable agents providing remember, recall, and reflect tools automatically, and offers both self-hosted (MIT licensed, free) and managed cloud options.
Vectorize pricing
Pricing model: Free
Self-hosted: Free - MIT open source license, single Docker command deployment, all four memory networks, Retain/Recall/Reflect APIs, MCP server built in, embedded PostgreSQL, community support via GitHub. Hindsight Cloud: Pay-as-you-go on token usage with no fixed monthly fee or seat pricing - starts free with free credits. Pricing per million tokens: Retain (store memories) $15.00, Recall (retrieve memories) $0.75, Reflect (synthesize patterns) $3.00, Iris Extract (structured extraction) $7.50, Mental Model Retrieve $0.25, Mental Model Refresh $3.00. Includes managed infrastructure, automatic scaling, daily backups, dashboard/analytics, team collaboration, support SLA available, 99.9% uptime guarantee. Enterprise: Custom deployment with bring-your-own-cloud or on-premises, dedicated infrastructure, custom SLA up to 99.95%, up to 24x7 support with 30-minute response, SSO/RBAC, custom integrations, onboarding/training - contact sales.
Vectorize pros
- 94.6% on LongMemEval benchmark - highest score of any memory system
- Open source and MIT licensed with no usage limits or restrictions
- Self-host free with a single Docker command
- Four parallel retrieval strategies for comprehensive memory recall
- Token budgets instead of top-K for predictable context and costs
- Learns from mistakes when tool calls fail or users correct agents
- Automatically detects patterns through reflection layer synthesis
- Per-user memory with isolated persistent context
- Cross-session persistence - context survives weeks later
- Fast memory recall under 100ms with parallel search
- Model-agnostic - works with any LLM without losing learned data
- Conflict detection preserves both states of changing facts with timestamps
- Entity resolution handles hundreds of entities in 3 queries (99.5% reduction)
- Temporal search with causal chains for time-aware reasoning
- SOC2 Type 2 certified for production security
- MCP server built in with remember, recall, reflect tools
- Python SDK and full REST API available
- No phone-home telemetry for self-hosted version
- Structured memory separates evidence from inference
- Hindsight Cloud offers 99.9% uptime guarantee with managed infrastructure
Vectorize cons
- Self-hosted requires managing your own infrastructure and Docker deployment
- Hindsight Cloud uses pay-as-you-go token pricing which can be unpredictable
- Retain operation costs $15 per million tokens - highest cost operation
- No fixed monthly fee means costs scale directly with usage volume
- Self-hosted version only has community support via GitHub
- Cloud tier has 12x5 email support only (not 24x7)
- 12-hour response SLA for Cloud may be too slow for production
- Enterprise plan requires contacting sales - no self-serve pricing
- Newer product with smaller community compared to established alternatives
- Reflection layer may add latency to agent operations
Frequently asked questions about Vectorize
Is Hindsight really free to self-host?
Yes. Hindsight is MIT licensed with no usage limits, no telemetry, and no restrictions. You can run it in production on your own infrastructure at no cost using a single Docker command.
What is the difference between self-hosted and Hindsight Cloud?
Self-hosted means you run Hindsight on your own servers using Docker and manage the infrastructure yourself. Hindsight Cloud is the fully managed service where Vectorize handles infrastructure, scaling, backups, and updates so you can focus on building your agent without ops work.
Can I migrate from self-hosted to Cloud later?
Yes. You can contact Vectorize for assistance migrating your memory data from a self-hosted instance to Hindsight Cloud. The same API and MCP interface are used for both options.
How does Hindsight Cloud pricing work?
Hindsight Cloud is pay-as-you-go based on token usage with no fixed monthly fee and no per-seat pricing. You pay for tokens consumed by each operation: Retain (storing memories), Recall (retrieving memories), Reflect (synthesizing patterns), Iris Extract (structured extraction), and Mental Model (retrieve and refresh operations). Free credits are provided to start so you can build and test before charges kick in.
How much do tokens cost for each operation?
All pricing is per million tokens: Retain (store memories) is $15.00, Recall (retrieve memories) is $0.75, Reflect (synthesize patterns) is $3.00, Iris Extract (structured extraction) is $7.50, Mental Model Retrieve is $0.25, and Mental Model Refresh is $3.00.
Do you offer volume discounts for high token usage?
Yes. If you expect significant token volume, you can contact Vectorize to discuss custom pricing and volume discounts.
Is there a free tier to start with?
You get free credits to get started, so you can build and test a real agent before any charges kick in. This allows you to evaluate the product before paying.
What benchmark scores does Hindsight achieve?
Hindsight achieves 94.6% on LongMemEval, the peer-reviewed benchmark for agent memory systems. This is the highest score of any memory system and is independently reproducible with open source benchmark code. It outperforms GPT-4o (60.2%), Zep (71.2%), and Supermemory (85.2%).
What retrieval strategies does Hindsight use?
Hindsight uses four retrieval strategies running in parallel: Dense vector search for semantic similarity via embeddings, Sparse vector search using BM25 for keyword matching, Graph traversal for entity relationship connections, and Temporal search for time-aware retrieval with causal chains. Results merge with token budgets rather than top-K for predictable context size and cost.
Which agent frameworks and LLMs does Hindsight work with?
Hindsight is model-agnostic and works with any LLM - you can swap LLMs without losing what your agent learned. It integrates with existing agent frameworks and is designed for MCP-capable agents, automatically providing remember, recall, and reflect tools. Python SDK and full REST API are available for integration.