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

← leaderboard

ADE Brain

episodes, cards, a fact layer that retires a value when a newer one contradicts it. https://github.com/adecubed/adebench

Note: the memory of adebench's author, and its code is private: nobody else can rerun this row. It runs on the same set and rules as every other memory

Core
89% · 48.8 / 55 (mean)
Core (public2)
public2 run in progress
Private set
after the public2 runs
Full
94% · 93.8 / 100 · 100 of 100 measured
Runs
5 · core 48.8–48.8 · the details below are the median run
Common with gbrain
83.8 vs 75.4 on 90
Chars per answer
1,239 · answer after 394
Write → visible
100 ms · p95 100 ms
Write-back
0 of 2 degraded answers came back
Reached through
REST
Model on write
yes: cards and distillation (Gemini)
Key needed
model provider
Where it runs
self-hosted
License
private
Deterministic
yes in five builds, model-written cards included
Configuration
VersionADE Brain service 1.0.0 (private code), evaluation instance
Modeself-hosted evaluation instance, empty before the import, reached over REST (voice door)
DoorPOST /sofia/ask {query, sources: [working, semantic, episodic, conversations, files], include_raw: true}, text as the voice client hears it: a latest-events block (from /memory/episodic/recent) around the summary, card first
Resultsthe Brain's own retrieval limits (not stated); latest-events block of up to 8 lines; cut at 2400 characters, the voice client's budget (adebench default)
Model insidegemini-3-flash-preview (Google): writes each entity's card from the facts it holds; episode distillation runs in the background and had not run on any of the 4 episodes when the benchmark read the memory
Embeddingssentence-transformers paraphrase-multilingual-MiniLM-L12-v2, 384 dimensions, local
StorageSQLite databases with full-text indexes
Not defaultdecay off (evaluation instance); BRAIN_LANG=en, the golden set's language; the set's repository indexed at the full-text stage only (folder summaries need the model and are not part of the set)

Notes: empty instance, BRAIN_LANG=en, cards and distillation with gemini-3-flash-preview, the set's repository indexed (full-text stage only). Run 2026-09-30, door voice, fingerprint 2942d1d84b. Report: examples/ade_synthetic_report/reference.md · All runs: examples/ade_synthetic_report/repeats.json

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 · 11 pass · 0 fail
updates
10.0
of 10 · 3 pass · 0 fail
time
10.0
of 10 · 6 pass · 0 fail
live state
10.0
of 10 · 7 pass · 0 fail
abstention
10.0
of 10 · 4 pass · 0 fail
file search
10.0
of 10 · 6 pass · 0 fail
graph
10.0
of 10 · 5 pass · 0 fail