The Cognitive Substrate for Persistent AI Agents Powered by Native HTAP
PhoebeDB makes the Cognitive Step a first-class database primitive — cross-type atomicity, snapshot-consistent reads, and causal replay in a single kernel.
⇩One modest dual-socket server. Over 23.5 Million TPM throughput. Driven by HammerDB's TPC-C workload. 20x higher than PostgreSQL 17.
What is PhoebeDB
PhoebeDB is the cognitive substrate for persistent AI agents. Its kernel treats the Cognitive Step — one complete agent reasoning cycle — as a first-class transaction primitive: relational state, vector memory, and trace data commit or roll back atomically. Built on a native HTAP engine with full PostgreSQL compatibility, PhoebeDB eliminates the fragmented stack that every production agent system eventually hits.
Key Features (Preview)
Cognitive Step API. The agent reasoning cycle is a first-class database primitive. State, vector memory, and trace data commit or roll back as one atomic unit via a clean SQL protocol — no application-layer coordination required.
Cross-type atomicity. The ANN index participates in the kernel's MVCC protocol as a first-class citizen. On rollback, relational writes, vector index updates, and trace entries are all discarded together. No ghost state. No partial results.
Snapshot-consistent reads within a step. All reads in a single cognitive step — relational queries, vector retrievals, event polls — see the same committed snapshot, enforced by the kernel. Agents in shared memory pools never reason from mismatched system states.
Causal replay with branch isolation. Every write carries a kernel-assigned step identifier. Reconstruct exactly what the agent knew at any prior step and test corrections against that exact historical state — without touching the live agent.
Native HTAP, no replication lag. PAX storage layout serves transactional writes and analytical queries on one physical copy. Cognitive management queries — contradiction detection, loop detection, memory health — run on live data in the same MVCC snapshot as concurrent writes.
Full PostgreSQL compatibility. PhoebeDB embraces the PostgreSQL ecosystem, aligning its kernel behavior with PostgreSQL, while enhancing plugin capabilities without altering existing interfaces. Learn more ...
Why choose PhoebeDB?
No Ghost Memory Entries
Cross-type atomicity means a failed cognitive step leaves no partial trace. The agent state update, vector memory write, and trace entry all roll back together — by construction, not by cleanup code.
Consistent World View per Step
Each reasoning cycle reads a single kernel-enforced snapshot. In multi-agent systems sharing a memory pool, agents never assemble beliefs from mismatched states taken milliseconds apart.
Full Causal Audit and Replay
Every cognitive step is replayable from its exact kernel-assigned snapshot. Audit agent behavior, correct past decisions, and branch-test fixes against historical state — without touching the live system.
One Kernel, No Fragmented Stack
Replace PostgreSQL + Pinecone + Redis + Kafka + LangSmith with a single system. One operational surface, one recovery boundary, one place where something can go wrong at 3am.
Trusted Ecosystem, Zero Porting Cost
PhoebeDB is 100% aligned with PostgreSQL behavior. Existing drivers, tooling, and plugins work without change — no migration cost, no learning curve for your team.
Demos
Hello World!
PhoebeDB's all-new kernel is seamlessly integrated with the latest PostgreSQL, ensuring that all extermal behaviors and interfaces remain fully aligned with standard PostgreSQL.
HammerDB TPC-C 23.5 Million TPM
On one modest dural-socket server, PhoebeDB sustained a transaction rate of over 23.6 milion TPM. Under identical conditions, PhoebeDB delivered an end-to-end throughput approximately 20x higher than PostareSQL 17.