Verging Labs
Karpathy Wiki

Karpathy Wiki

Agentic Memory

Where it ranks

Hosted providers index what you write on their own infrastructure; local stores do not, so the two are also ranked apart. Agentic Memory Index v0.1, released 2026-08-03.

Time to ready
local
#1 of all measured tools (tied)
Agentic Memory Index score
98.5
#1 of all measured tools
Overall pass rate
98%
#1 of all measured tools
Recall speed, median
0.03s
#1 of all measured tools
Store speed, median
0.05s
#1 of all measured tools
Answer latency, median
2.7s
#1 of all measured tools
Cost per 1,000 successful answers
$568.93
#8 of all measured tools

Podium

1st · Agentic Memory Index
98.5
1st · Direct Recall
98%
1st · Updated Facts
100%

Agentic Memory Index

v0.1
98.5
CI 96.7–99.4
Rank 1 of 8
Direct Recall
98%
Updated Facts
100%
Thread Growth
94%
Synthesis
100%
Long Term
99%
False memory check
99%
Methodology →

Failure Attributioni

Question Not Addressed
1.0%
Incorrect Answer
1.0%
Outdated Answer
0.0%
Fabricated Memory
0.0%

Long-term retentioni

How well early facts survive: probed at five checkpoints across 56 working sessions.

20%40%60%80%100% 1021334556working sessions elapsed C
20%40%60%80%100% 1021334556working sessions elapsed C

Speed

Onboard
1.4 min
Time to ready
local
Time to store
62.0s
Answer speed
2.7s

Cost

Per 1,000 successful answers: agent token spend, the entry paid tier amortized at its capacity, and embeddings where billed separately.

Cost per 1k successful answers
$568.93
Agent Tokens
$568.93
Subscription
$0.00
Embeddings
$0.00

Pricing

Free tier
Always
Starting at
$0
Open source
Open method

List prices as of Jul 27, 2026.