Readback SI, Super Intelligence systems

Agents that remember, cite their source, and say when they don’t know.

Your agents remember what happened, show you where every answer came from, and say so when nothing in the record supports one. Built on LBrain: open source, local first, no API key.

Start building your agentic team Read the research

Same question. Same model. One has the record.

Pick a question. On the left, a model answers from memory alone. On the right, the same model answers through LBrain.

Without a record

With LBrain

8 architectures, 7 organisations, 168 sealed probes. For the failure we tested, changing the model did not remove made-up answers. Changing the record it was served moved them by 34.4 percentage points. The Record, Not the Model
24 frontier models, questions with no knowable answer. 22 declined to answer. The two that invented answers did so 56% and 62% of the time without a record, and 3% with one. Grounded Memory Makes Models Faithful, Not Truthful

The building blocks

Four systems we run our own company on, every day. Each is open to inspect.

Governed memory

Every answer carries its record, its date and its source. A superseded record stays findable, but it never wins.

See it work

Epochs

The memory is rebuilt, validated and published in one atomic step. A build that fails validation is never served.

See it work

Seats with identity

Named agent seats, each with a permanent gcx:// address and a profile inscribed on Arweave that anyone can verify.

See it work

The review gate

Deterministic code decides what the model may claim. After its context is compacted, an agent's first action waits on the record, not the summary.

See it work

Start in three commands

LBrain runs on your machine. Embeddings are computed on the device, so your records never leave it.

Open source under BSD-3-Clause. Works with any agent that speaks MCP.

Installpip install lbrain
Give it your recordslbrain import ./your-notes
Ask, and see the sourcelbrain query "what did we decide?"

The research behind it

Preprints with DOIs on Zenodo. Each states what was tested and what was not.

Preprint, 2026-08-24The Record, Not the Model: Cross-Architecture Evidence that Near-Domain Retrieval Confabulation Is Dominated by the Served Record Preprint, 2026-10-10Grounded Memory Makes Models Faithful, Not Truthful: Recall, Abstention, and Poison-Relay across Memory Regimes and 24 Frontier Models Preprint, 2026-10-07The Relocatable Context Window: Training-Free KV Surgery with an Accountable External Store Theory, 2026-06-14The Ancestry of Belief: Epistemic Lineage as a First-Class Primitive for Bounded Intelligences, Human and Machine