LoCoMo v1, all 1,986 questions, Top 10Season 1 · Rank 7 of 7

Zep Graphiti open source

Zep Graphiti open source (in-process, local FalkorDB)zep/graphiti

Run history
  1. 1 Nov 2026, 15:18 UTC Started
  2. 1 Nov 2026, 16:19 UTC Stopped: retrieval failed (the memory system or the retrieval job)
  3. 1 Nov 2026, 17:18 UTC Resumed
  4. 1 Nov 2026, 19:38 UTC Finished

Recorded by the runner. Completed conversations were kept on resume; unfinished ones were redone.

Report a problem with this result (see the corrections policy)

The judge ruled the retrieved context correct for 978 of 1,986 questions (49.2%, 95% CI 46.4 to 52.0), using 426 tokens per question on average.

Figures

Context

Tokens per question, mean
426
Tokens, 95th percentile
732
Accuracy per 1,000 tokens
115.5
Contexts cut to the budget
0.0%

Speed

Retrieval, median
1.60 s
Retrieval, 95th percentile
2.34 s
Ingestion, total
6 h 26 min (all conversations added up)
Ingestion, per turn
3.94 s

Judge cost

Top 10
$10.10
Top 30
$15.01
Whole run
$25.10from price table 2026-10-04

Checks

Abstention accuracy
94.4%
No verdict (call failed)
0
Flagged by the judge
8

Where it fails

Accuracy by question category, with 95% confidence intervals. A low category is where the memory system misses the context the question needs.

Accuracy by category
CategoryCorrectAccuracy95% CITokensNo verdict
adversarial421 of 44694.4%92.8 to 96.24110
multi-hop75 of 28226.6%20.6 to 33.04340
open-domain34 of 9635.4%21.6 to 51.94280
single-hop320 of 84138.0%34.6 to 41.64290
temporal128 of 32139.9%31.5 to 47.44360

Every question

One square per question, grouped by category. Select a square to read the question, the gold answer, the judge's reason and the context it was given.

Loading questions…

Retrieval latency

The share of questions whose context came back within a given time. Further left is faster.

Judge

A season is one fixed set of judging rules: one pinned judge model, its settings and its prompt rule on every question. When the rules change a new season starts, and earlier runs are judged again from their stored contexts, without re-running the memory system. Seasons

Model
claude-sonnet-5-5
Route
anthropic
Season
Season 1
Effort
medium
Output limit
4,000 tokens
Prompt
v3 sha256:ce6d9750331c79d1bb3a0737109c458fef4ba9fd8cf929c75c1c1cf9407a4b78
Served by
claude-sonnet-5-5

How this result was produced

Everything below is part of the signed result. Change any of it and the signature no longer matches.

Run

Run id
20261101T151858Z-33a8ed
Mode
hosted runner
Finished
1 Nov 2026
Duration
4 h 19 min
memjudge version
0.1.0
Commit
a73cf373d958
Image
sha256:8119890e8697d791a388f2397208dbc8cb236a01319e1d74e36fd616660fce2c
Context budget
8,192 tokens (o200k_base)
Conversations at a time
4
Submitter affiliated
no

Dataset

Name
locomo v1
Questions
1,986
Licence
CC-BY-NC-4.0
Corrections applied
0
Hash
sha256:9c7812f2f45ce12b06cd9d3fc55736c3a8a83d2b263c11f28c292c475f6ddf38
Source file hash
79fa87e90f04081343b8c8debecb80a9a6842b76a7aa537dc9fdf651ea698ff4

Connector

Connector
zep/graphiti
Written by
MemJudge maintainers
SDK
graphiti-core[falkordb]==0.31.1
Code hash
sha256:d1a55436d110c08a023de10be6bd4d4af17998d8249e65770d42fb63f9f1db2b
Models
gpt-6-luna, text-embedding-3-small
llm model
gpt-6-luna
small model
gpt-6-luna
embedder model
text-embedding-3-small
falkordb host
localhost
falkordb port
6379

Raw files

FileSHA-256
conversations.jsonl5cead1b3569f89146e22de6a290c44c83e3b6826c2489195c810495c7f528bcc
questions-top10.jsonl007375471476b4d61feeb039fb4e86a029416d29c5f2ead7538e97680b807b57
questions-top30.jsonl83a60cb03d53d458dd6986aa6a9e7707941139b513000edeb02c5acba7d27b3a

Check it yourself

The signed result is signed with key 478df5eb628bfa74, one of the trusted keys published in the leaderboard repository. Download it and verify it with the memjudge command-line tool; add --artifactswith the raw files to check those too. How signing works.

memjudge verify 20261101T151858Z-33a8ed.json

Download the signed result

Questions, answers and conversation excerpts on this page come from LoCoMo (Maharana et al., 2024, Snap Research) and are used under CC BY-NC 4.0 (non-commercial use only). They are reformatted, and corrected where MemJudge's errata apply. MemJudge's own verdicts and scores are CC BY 4.0.