Agent memory infrastructure

Memory that knows what changed.

UrBrain is a hosted memory layer for agent systems and RAG-backed products. It stores source documents, document versions, memories, embeddings, and retrieval policy under account, organization, and workspace boundaries — so agents retrieve current truth, scoped to the people and teams allowed to see it.

Ingest at a glance

New version in. Only what changed gets re-embedded.

A document update arrives. UrBrain finds the prior version it already stores and hands both to StructuredMerge, which returns the chunks that actually changed. UrBrain re-embeds only those chunks and updates the workspace index. Everything else is left alone — saving compute, embedding tokens, and review time.

UrBrain ingest flow with selective re-embeddingNew document versionarrives via upload, API, or connectorPrior version (UrBrain store)canonical record + content hashStructuredMergedeterministic delta · stable chunk IDs · content hashesprior +new bytesunchangedchunks skippedonly changed chunksUrBrain re-embeds and updates the indextombstones removed chunks · upserts changed chunks · skips the restworkspace memory current for retrieval

What UrBrain is

A durable second brain for agent systems.

Most agent stacks treat memory as a side effect of a chat transcript: vectors are written when a user says something interesting, and retrieved when a prompt looks similar. That model breaks the moment a document changes, a person leaves a workspace, a policy is rewritten, or a customer asks for an audit trail.

UrBrain takes the opposite shape. Source documents are stored as durable versions. Updates are processed through deterministic deltas that produce upserts and tombstones, so memory tracks the current state of each source. Every memory carries first-class source identity, lifecycle state, and workspace scope, and retrieval honors all three before returning a result to an agent.

The product is hosted at ur-brain.tech. The same vertical stack — API, workers, internal ingest service, PostgreSQL with pgvector, and object storage — can be deployed as an isolated stack for customers whose documents cannot share infrastructure with anyone else.

Where it fits

One product boundary for the memory layer.

UrBrain owns the customer relationship, the canonical document record, and retrieval. StructuredMerge runs as an internal service to keep document updates stable and reviewable. Customers subscribe to UrBrain and never have to manage a second product to get version-aware document memory.

01 Documents and versions
The canonical raw payload, version history, content hashes, and source connector metadata live inside UrBrain. Every change becomes a durable version that the rest of the pipeline can reason about.
02 Memories and embeddings
Promoted chunks become first-class memories with stable source IDs, content hashes, and lifecycle state managed by the product. Embeddings live next to the policy that gates them.
03 Retrieval and policy
Hybrid search over pgvector cosine similarity and PostgreSQL trigram lookup, fused with reciprocal rank fusion and filtered by account, organization, workspace, memory type, and lifecycle.
04 Access surfaces
Hosted MCP for agents, a REST API for applications, and runtime adapters for the agent stacks teams already use. Connectors are designed as portability surfaces, not one-way importers.
Configure UrBrain once, then let local agents remember and retrieve in the background — without making users manage each runtime's private memory system.

— UrBrain strategy

Built by 9thBit

Talk to the team.

UrBrain is a 9thBit, LLC project. Reach the team directly to talk about pilot deployments, isolated stacks, or runtime adapter work.