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

← leaderboard

cognee

knowledge graph plus vector index built from what is added (cognify). https://github.com/topoteretes/cognee

Core
63% · 34.4 / 55 (mean)
Core (public2)
public2 run in progress
Private set
after the public2 runs
Full
68% · 44.4 / 65 · 65 of 100 measured
Runs
5 · core 33.7–35.4 · the details below are the median run
Common with gbrain
44.6 vs 50.4 on 65
Chars per answer
4,781 · answer after 297
Write → visible
921 ms · p95 1.0 s
Write-back
2 of 2 degraded answers came back
Reached through
Python library
Model on write
yes: graph extraction (Gemini)
Key needed
LLM provider
Where it runs
local
License
Apache-2.0
Deterministic
no: extraction is a model call
Configuration
Versioncognee 1.6.1
ModePython library in its LLM-based (Gemini) mode, run through a JSON-lines bridge in cognee's own Python
Doorcognee.recall(question, datasets=["adebench"], only_context=True, retriever_specific_config={include_external_metadata: true, external_metadata_keys: ["date"]}, session_id=<fresh>), no query_type (cognee routes to HYBRID_COMPLETION)
Resultsno top_k sent (cognee's own default, value not stated), no cut
Model insidegemini-3-flash-preview (Google): graph extraction (entities, relations) and summaries at cognify; no answer model on this door
Embeddingsgemini-embedding-001, 768 dimensions (Google)
StorageLanceDB (vectors), Ladybug (graph), SQLite (metadata)
Not defaultinclude_external_metadata with key 'date' (opt-in: puts each passage's date in the context); a fresh session_id per recall (session memory is on by default and would feed earlier questions back); telemetry off (reason not stated)

Notes: 1.6.1, gemini-3-flash-preview and gemini-embedding-001, door = recall context (only_context), dates shown through include_external_metadata, fresh session per call. Run 2026-09-30, door context, fingerprint 9c9c8684e6. Report: examples/cognee_report/reference.md · All runs: examples/cognee_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
not measuredsection not measured: no entity cards in this memory
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 841 ms; STALE: the old value is still delivered (17 reads with both)
FAIL restating the current value does not pile up a second copy — 5 copies delivered
time
8.3
of 10 · 5 pass · 1 fail
FAIL semantic memories carry their age (55/70) — 79%
live state
10.0
of 10 · 7 pass · 0 fail
abstention
4.2
of 10 · 0 pass · 4 fail
FAIL What is the Zarpetta module? — 10 keyword facts for something that does not exist
FAIL Who is Ottavio Brambillesco? — episodes presented as direct matches; 9 keyword facts for something that does not exist
FAIL Which port does the Fulmicotone service use? — episodes presented as direct matches; 9 keyword facts for something that does not exist
FAIL What does the calendar plugin Girandola do? — episodes presented as direct matches; 9 keyword facts for something that does not exist
file search
not measuredsection not measured: no repo configured (--repo)
graph
not measuredsection not measured: no case with evidence (all SKIP or no cases)