Self-hostable · open source (Apache-2.0) · built on Graphiti × LlamaIndex

The auditable, self-hostable belief ledger for agents.

Plain RAG can tell you what's true now. RAGBrain tells you what your agent believed at any past moment - and proves it. Temporally-correct context and cited answers, running in your own VPC, where your data never leaves.

See the live head-to-head → Read the code

The one thing plain RAG can't do

Ask "as of when?"

Modern RAG (and hosted memory) has no time axis. It answers from whatever it retrieves today. RAGBrain filters context to the moment you ask about, so the answer is temporally correct - and cites the fact it stood on.

Same question, add "as of 2015"	Answer
Plain RAG	Can't isolate the past - no time axis. It sees both the old and new fact and hedges or picks wrong.
RAGBrain (as of 2015)

Watch it in the live demo →

Positioned by contrast

We are the thing the recall-optimized clouds vacated

Not a worse hosted-memory service. The answer to a question their governed cloud structurally can't answer - for the regulated buyer who can't send data out. We tie on retrieval (we inherit Graphiti's); we win where they're blank.

Capability	RAGBrain	Hosted recall-optimized memory
Standard retrieval / recall	Tie - inherited, not a win	Strong
Temporal correctness / as-of replay	Yes	- no time axis
Auditable provenance - "why did this change?"	Yes	-
System-time replay (what did it know then)	Yes	-
Self-hostable / in-VPC / data never leaves	Yes	- hosted SaaS

The tie row is the point: an all-green matrix is the one nobody believes.

Real benchmark - every number from a live run

Where we tie, and where it isn't close

Same corpus, same model, same embeddings. The only difference is memory. Panel 1 is an honest tie on stable facts. Panel 2 is what a temporal store is for.

Honesty is the marketing

What we are - and what we're not

Structured input is deterministic. Facts you assert (OKF fact keys) get precise temporal validity. Facts extracted from raw prose are only as good as the extraction LLM - each served fact is labeled with its confidence, so nothing is laundered.
Retrieval is inherited, not class-leading. We use Graphiti's retrieval; we don't out-retrieve the recall specialists. An optional reranker is measured, off by default, on by evidence.
Not "first" at temporal RAG. Validity filters and temporal-graph RAG exist. What we ship is honest system-time replay - the un-knowing invariant - as auditable, self-hostable infrastructure.
Read-only audit surface. The replay dashboard is a window onto the ledger, never a second product; the answer path never writes.

Try it

Run it in your own environment

Your infrastructure, your LLM, your embedder - your data never leaves. For an audit product, the strongest CTA is: read the code that produces the audit trail.

Read the code → Reproduce the demo

Reproduce every number here: python demo/capture_demo.py and python demo/benchmark.py.