Tacnode Context Lake
by Tacnode
PostgreSQL-compatible context lake giving AI agents shared, live, transactional state at decision time
Tacnode Context Lake is a PostgreSQL-compatible data platform that gives AI agents and automated services one shared, continuously updated view of enterprise state. It holds transactional rows, JSON, text and vector embeddings in a single ACID system with sub-second freshness, so fraud, personalisation and underwriting agents all decide against the same committed snapshot instead of three stale copies.
Tacnode Context Lake is a distributed data platform built around what the company calls the context gap: automated systems making decisions in milliseconds against data that is minutes or hours old and scattered across an OLTP replica, a feature store and a vector database. It collapses those into one transactional system holding structured rows, semi-structured JSON, unstructured text and vector embeddings together, and serves them with sub-second ingestion-to-query freshness under high concurrency. It exposes the PostgreSQL wire protocol natively rather than through a translation layer, so psql, any ORM, existing drivers and BI tools connect unchanged, and full SQL — joins, CTEs, window functions — works alongside native HNSW and IVFFlat vector indexes, inverted, JSON and bitmap indexes, and incremental materialized views that keep derived features current without batch jobs. Compute and storage scale independently with multi-tenant workload isolation and multi-AZ failover; isolation levels include serializable, snapshot and read committed; and all state is versioned over time, giving time-travel queries and safe replay as a first-class capability rather than bolt-on tooling. Semantic Operators expose LLM calls such as llm_classify, llm_extract and llm_summarize directly inside SQL. Tacnode emerged from stealth in January 2026 under CEO and chief architect Xiaowei Jiang, previously a database and distributed-systems engineer at Microsoft, Meta and Alibaba, and sells directly and through the AWS Marketplace AI Agents and Tools category with Model Context Protocol support. DoorDash runs it in production for in-session personalisation, reporting response times cut from several minutes to hundreds of milliseconds. The company positions OLAP warehouses — Snowflake, ClickHouse, BigQuery, Redshift — as complementary rather than replaced: those serve analysts reviewing history, Tacnode serves agents deciding now. It carries SOC 2 Type II and HIPAA compliance and publishes per-node-hour pricing instead of routing every buyer to sales.
The platform-data or ML-infrastructure lead running real-time decisioning — fraud, pricing, personalisation, underwriting — who is currently gluing a feature store, a vector database and an OLTP read replica together and paying for the drift between them.
One PostgreSQL endpoint serves fresh operational state, computed features and vector search inside a single transaction, so concurrent agents stop making conflicting decisions from disagreeing copies of the same customer.
At a Glance
- Category
- Data & Analytics
- Pricing
- Usage-based, Contact for pricing
- Target Market
- CTOs, Data Engineers, ML Engineers, Platform Engineering Leaders, Enterprise Developers
- Deployment
- Cloud-first, Hybrid
Key Features
- ✓Native PostgreSQL wire protocol
Compatibility is a design constraint, not a shim, so existing psql clients, ORMs, drivers and BI tools connect without modification.
- ✓Unified multi-modal storage
Rows, JSON, unstructured text and vector embeddings live in one system with ACID transactions spanning all of them.
- ✓Native vector search
HNSW and IVFFlat indexes run in-database, removing the separate vector store and the sync pipeline that keeps drifting.
- ✓Semantic Operators
llm_classify, llm_extract and llm_summarize are callable directly from SQL, so LLM-derived signals compute where the data already sits.
- ✓Temporal versioning and time travel
All state is versioned over time as a core capability, enabling time-aware decisions and safe replay without external tooling.
- ✓Compute-storage separation
Storage and compute scale independently with workload isolation, so concurrent agent traffic and heavy queries do not interfere.
- ✓Incremental materialized views
Derived aggregates and features stay continuously current instead of being rebuilt on a batch schedule.
Capabilities
Use Cases
- •Fraud detection at transaction rate
Score transactions against the live stream with sub-millisecond decision latency instead of waiting for a batch load into a warehouse.
- •In-session personalisation
DoorDash runs Tacnode in production for in-session personalisation, cutting response times from several minutes to hundreds of milliseconds.
- •Credit and underwriting decisions
Evaluate an applicant against current obligations and behaviour rather than a snapshot that aged out hours ago.
- •Identity and policy enforcement
Enforce access and policy decisions consistently across services that would otherwise each read their own stale replica.
- •Shared agent memory
Externalise agent state into transactional infrastructure so multiple concurrent agents read one committed view of the world.
Ideal For
Best For
- ✓Real-time fraud decisioning where the agent must see the live transaction stream, not a batch-loaded snapshot
- ✓Serving ML features computed on current operational state to eliminate training/serving skew without a separate feature store
- ✓Shared agent memory that must be transactional and consistent across many concurrent agents
- ✓Consolidating a read replica, feature store and vector database into one PostgreSQL-compatible system
- ✓Low-latency decision-time aggregations over continuously updating state, such as credit and underwriting checks
Not Ideal For
- ✗Teams that require auditable open-source infrastructure — Tacnode is closed source, and an independent comparison of agent memory systems noted that every ACID, consistency and performance claim rests on documentation rather than inspectable code
- ✗Classic BI and historical reporting for human analysts; Tacnode itself frames Snowflake, BigQuery and Redshift as complements you keep running alongside it
- ✗Small teams and side projects — billing starts at a per-node hourly rate designed for always-on production concurrency, not intermittent use
- ✗Buyers who need the knowledge layer itself: Tacnode deliberately does not do entity extraction, relationship discovery or episodic memory the way Mem0, Letta or Cognee do
Deployment
Market Analysis
Pros
- ✓PostgreSQL compatibility means existing clients, ORMs, drivers and BI tools work unchanged — the cheapest possible adoption path
- ✓Collapses feature store, vector database and read replica into one system, deleting the sync code and the drift it causes
- ✓Published per-node-hour pricing with a calculator, which is unusual for enterprise data infrastructure
- ✓A named production reference at DoorDash with a concrete latency result (minutes to hundreds of milliseconds)
- ✓SOC 2 Type II and HIPAA compliance already in place, plus a BYOC option for regulated deployments
Cons
- ✗Closed source — an independent agent-memory comparison flagged that every ACID guarantee and performance claim is documentation-based rather than verifiable in code
- ✗Almost no independent practitioner discussion exists; the Hacker News mentions are largely from someone affiliated with the company, and there are no G2, Capterra or TrustRadius reviews
- ✗The vendor declines to disclose funding or investors, which is a real diligence gap for infrastructure that sits in the decision path
- ✗Not a warehouse replacement — you keep paying for Snowflake or BigQuery alongside it, so this is an added line item, not a consolidation saving
- ✗Time-travel retention defaults to roughly a 23-hour window according to third-party analysis, so long-horizon replay needs deliberate configuration
- ✗Deliberately not a knowledge layer: no entity extraction, relationship discovery or episodic memory, so an agent-memory team still needs something on top
Pricing
Commercial Edition
From $2.00/hour
- ✓Tacnode-hosted
- ✓Compute-storage separation
- ✓Advanced indexing and incremental materialized views
- ✓Point-in-time recovery
- ✓12x5 support
Business Critical Edition
From $3.00/hour
- ✓Everything in Commercial
- ✓Multiple workload isolation
- ✓Multi-AZ failover
- ✓Audit logging, data masking and encryption
- ✓24x7 support
Enterprise (BYOC)
Contact for pricing
- ✓Everything in Business Critical
- ✓Bring your own cloud, deployed in your VPC
- ✓Fully isolated environment with physical tenant separation
- ✓Dedicated support
Metered per node-hour rather than per seat: Commercial is $2.00/hour and Business Critical $3.00/hour, with tiered storage billed on compressed data at $0.1460/GB/month hot, $0.1230/GB/month cache and $0.0230/GB/month cold. Only the BYOC Enterprise edition is quote-based, so unlike most vendors in this category a workload can be costed from the public calculator before anyone talks to sales. No free tier or free trial is advertised.
Security & Compliance
Connect
Sources
This page was written from 6 sources, 3 on domains other than tacnode.io.
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Reducto
Agentic document parsing API that turns complex PDFs, scans and spreadsheets into citation-grounded data for AI, now with the 1¢-per-page r-1 model
Euno
AI-native context platform that gives enterprise agents governed, trustworthy data
Prevalent AI
A sovereign enterprise knowledge graph that gives security teams and AI agents context
StitcherAI
IT finance intelligence that puts cost context where the spending decision happens