# Levirge Brain vs GBrain

[GBrain](https://github.com/garrytan/gbrain), from Garry Tan, is an
MIT-licensed memory server for agents, also sold hosted as
[gbrain.io](https://gbrain.io/). Brain is shared memory for teams of agents
and people. Both give every MCP client the same memory. They differ on who
writes the answer. GBrain details come from its repo and site, reviewed in Aug
and Oct 2026.

## Where GBrain is ahead

| Capability | GBrain | Brain |
| --- | --- | --- |
| Synthesised answers | `think` writes a cited answer and notes what the brain doesn't know. | Returns cited evidence plus a coverage note (no match, stale, thin entity). Your agent writes the answer. |
| Connectors | Gmail, Calendar, Contacts, voice calls and meeting webhooks feed it directly. | Agents, the API, the web UI and Obsidian write to it. |
| Code intelligence | Callers, callees, definitions, references and blast radius over indexed code. | None. |
| Open source and self-hosting | MIT. Runs on a laptop; can run fully on local models (Ollama, llama.cpp). | Hosted. On-premise is not generally available yet. |
| Published evals | LongMemEval-S and its own BrainBench, with reproduction commands. | Not published. |

## Both do

- **Hybrid retrieval:** keyword and semantic results fused (RRF), then a
  cross-encoder rerank.
- **Entity graph:** typed links between people, companies and topics.
- **Corrections:** a replaced fact is superseded, not silently overwritten.
- **Near-duplicate checks** when knowledge is written.
- **Secret scanning** at ingest.
- **Skills** stored beside the knowledge.
- **Shared access:** GBrain grants by source and operation; Brain grants by
  vault.

## Only in Brain

| Capability | Brain | GBrain |
| --- | --- | --- |
| Agent-to-agent handoffs | Durable inbox between agents. | None. |
| Lineage and trust | Facts trace to source documents, with trust tier and origin; runbooks are draft or ratified. | Provenance and a source-tier ranking boost; no ratification step. |
| Work continuity | Sessions, todos and reminders next to the memory. | Memory, skills and jobs; no todos or reminders. |
| Tool surface | About 20 MCP tools. | 151 tools; a ~40-tool starter surface for clients that cap tools. |

## Different by design

- **Who writes the answer.** GBrain puts an LLM after retrieval to compose
  the answer. Brain keeps LLMs out of retrieval: results are fast,
  deterministic evidence, and the agent holding the task context does the
  reasoning.
- **When curation runs.** GBrain curates in a nightly cycle of background
  jobs. Brain curates when knowledge is written, and records each decision
  with its provenance.

## Where each fits

- **GBrain:** one person or company that wants a brain answering for them,
  fed from email, calendar and calls, free to self-host.
- **Brain:** several agents and people working from the same memory, passing
  work between agents, with facts traced to their source.

## Benchmarks

GBrain reports LongMemEval-S recall_all@5 of 95.53% (93.40% with rerank off),
retrieval only, measured 6 Sep 2026 in
[its docs](https://github.com/garrytan/gbrain). Brain hasn't published scores
under a third-party harness, so we don't put the two side by side.

## Method

GBrain: repo `garrytan/gbrain` reviewed 7 Aug 2026, re-checked at commit
`f3f9dbd` on 20 Oct 2026; gbrain.io read 20 Oct 2026. Brain: checked against
its source on 20 Oct 2026. Both projects move fast; check the current docs.