©2026 adebench · adecubedbuilt 2026-11-04 from the reports in the repo

← leaderboard

gbrain

entity pages, chronicle, remember/recall/forget, hybrid search. https://github.com/garrytan/gbrain

Core
73% · 40.4 / 55 (mean)
Core (public2)
78% · 43.1 / 55 (mean)
Private set
lower by 5.8 core points than the set explains · δ 3 · shift −2.1 from 5 other memories · public2 3 builds, private 1 · holdout-v1 3cd42b9f
Full
84% · 75.4 / 90 · 90 of 100 measured
Runs
3 · core 40.4–40.4 · the details below are the median run
Common with gbrain
reference
Chars per answer
1,321 · answer after 162
Write → visible
2.3 s · p95 2.4 s
Write-back
2 of 2 degraded answers came back
Reached through
CLI / MCP
Model on write
embeddings only
Key needed
none (local Ollama)
Where it runs
local
License
MIT
Deterministic
yes
Configuration
Versiongbrain 0.50.0.0
Modelocal, every operation through `gbrain call`, the local dispatch of its MCP tools
Doorsearch(query, limit=8, snippet_chars=0) plus recall(query, limit=8), plus get_page for the entity pages the question names
Resultslimit 8 on search and recall (adapter's choice; gbrain's defaults not stated), no character cut; when every hit is a weak semantic match only two are kept
Model insidenone
Embeddingsollama:nomic-embed-text, 768 dimensions, local Ollama
StoragePGLite
Not defaultembedding model ollama:nomic-embed-text (768) set at init (without an embedding provider gbrain has keyword search only)

Notes: gbrain 0.50, PGLite, embeddings ollama:nomic-embed-text (768), door search, no character cut. Run 2026-09-30, door search, fingerprint 5c9ffcacad. Report: examples/gbrain_report/reference.md · All runs: examples/gbrain_report/repeats.json · public2: examples/gbrain_report/reference_public2.md

Sections
door
18.8
of 25 · 6 pass · 2 fail
FAIL Which version of MailBridge is installed for the mailbox? — STALE value delivered next to the current one: 1.3.0
FAIL What is the owner's phone number? — missing +39
cards
15.0
of 15 · 8 pass · 0 fail
updates
3.3
of 10 · 1 pass · 2 fail
FAIL a new write with a changed value replaces the old one at the door (no id given) — new value served in 2374 ms; STALE: the old value is still delivered (10 reads with both)
FAIL restating the current value does not pile up a second copy — 2 copies delivered
time
10.0
of 10 · 5 pass · 0 fail
live state
10.0
of 10 · 7 pass · 0 fail
abstention
8.3
of 10 · 3 pass · 1 fail
FAIL What does the calendar plugin Girandola do? — produced an entity card; 8 keyword facts for something that does not exist
file search
not measuredsection not measured: no repo configured (--repo)
graph
10.0
of 10 · 4 pass · 0 fail