Original analysis

Measurements we ran on our own machine, published with the method and the limits next to the numbers.

Articles and the weekly digest share one feed at /rss/blog.

Agent memory and context2026-10-25 15:50

ai-memory vs the memory Hermes already runs: 837 items, 24 probes

Same 837 items and 24 probes. ai-memory is correct in the top 5 for 18 of 24 at 290.8 ms. The injected memory block plus search_files answers 20 of 24 at 0.1 ms.

Read the measurements

akitaonrails/ai-memory

Agent memory and context

How much an agent remembers, how much context survives a tool call, and what each memory layer costs.

Output and token budget

What a tool actually removes from the bytes an agent reads — the ceiling, the median, and the cases where it removes nothing.

Agent tooling and code

The tools around the agent: code generation, code understanding, sandboxes, sessions, databases, and job-hunt automation.

Scraping and reach

Getting pages that do not want to be fetched: which fetcher opened a page, which one quietly returned nothing.

Media and rendering

Producing and fetching media on one machine: render throughput, file hashes across renders, and downloaders on the same URLs.

ai-memory vs the memory Hermes already runs: 837 items, 24 probes

Agent memory and context

headroom saves 34.5% on a tool result and 0% on the same bytes in a user message

claude-mem says 10x, our memory says 3.3x: pricing a retrieval claim on 837 real items

Agent memory, three ways: hindsight vs HydraDB vs keyword search on our own 837 items

MemPalace: 96.6% R@5 on LongMemEval measured at 96.0% on its own 50-question split, with the network switched off

headroom: the 10,144 → 1,260 demo reproduces at 86% on our worker logs, and at 2.8% through the shipped pipeline

Hindsight: we installed the 2.89 MB memory binary and checked whether it remembers

context-mode: we tested the 98% context claim with our own payloads

Output and token budget

Nine skills, one React page: what each skill cost in tokens

i-have-adhd: 10 output rules, measured on the agent they were not written for

OmniRoute: the 1.62B free-token headline recomputes to the digit, and 20 of 20 channels wanted a key first

caveman: the 129.8x page number is one filtered answer, and our lockfile came back untouched

rtk cuts 92% off ls -R and 0% off cat: the ceiling, the median, and the two commands it leaves alone

Agent tooling and code

graphify: 100% edge accuracy, 73% query accuracy, 1.7x more tokens than grep

career-ops on 8 real postings: the "still open?" check was 8 for 8, the "does it fit?" half only reached a keyword list

FalkorDB on our own dependency graph: five questions, three engines, and a load that lied twice

e2e: seven replays, no model calls, and the 25 seconds a renamed button costs

strix never finished one run, and our hand-run probes produced their own false positives

dsh free models: 7 of 13 answer, and the two in the headline live on another host

herdr: the work survives the client, and a restart brings back only the layout

OpenShell: 22 of 32 escapes refused, and the path its own docs say a policy cannot expose

magpie: 1.26 ms per request through the local gateway, and 1 of 4 agent configs comes back byte-identical

ponytail: 66% less code on a real repo, and the one task where it wrote nothing at all

DBX as an agent database layer: we measured the query path and the write gates

Is DBX really a 25 MB database client? We measured it on a Mac

Scraping and reach

firecrawl vs Scrapling vs Crawl4AI on the same ten pages

caveman vs rtk vs Scrapling on the same targets: what each one shrinks

Agent Reach: 2 of 5 targets read with no key, and the Reddit route its own docs rule out

Scrapling: three fetchers, 900 MB of install, and one Cloudflare page that opened

We ran Moli and headless Chrome side by side on a Mac, here is what we measured

Media and rendering

yoinks: 6 of 8 URLs, +48.7 s per file, and a picker that says mp4 when the file is AVI

hyperframes: the same 27-second clip came out as two different files in four renders, and neither of them took 4 seconds

fframes: 27.4 seconds of video in 4.3 seconds, and the encoder fallback that fooled us first

ai-memory vs the memory Hermes already runs: 837 items, 24 probes

Agent memory and context

headroom saves 34.5% on a tool result and 0% on the same bytes in a user message

claude-mem says 10x, our memory says 3.3x: pricing a retrieval claim on 837 real items

Agent memory, three ways: hindsight vs HydraDB vs keyword search on our own 837 items

MemPalace: 96.6% R@5 on LongMemEval measured at 96.0% on its own 50-question split, with the network switched off

headroom: the 10,144 → 1,260 demo reproduces at 86% on our worker logs, and at 2.8% through the shipped pipeline

Hindsight: we installed the 2.89 MB memory binary and checked whether it remembers

context-mode: we tested the 98% context claim with our own payloads

Output and token budget

Nine skills, one React page: what each skill cost in tokens

i-have-adhd: 10 output rules, measured on the agent they were not written for

OmniRoute: the 1.62B free-token headline recomputes to the digit, and 20 of 20 channels wanted a key first

caveman: the 129.8x page number is one filtered answer, and our lockfile came back untouched

rtk cuts 92% off ls -R and 0% off cat: the ceiling, the median, and the two commands it leaves alone

Agent tooling and code

graphify: 100% edge accuracy, 73% query accuracy, 1.7x more tokens than grep

career-ops on 8 real postings: the "still open?" check was 8 for 8, the "does it fit?" half only reached a keyword list

FalkorDB on our own dependency graph: five questions, three engines, and a load that lied twice

e2e: seven replays, no model calls, and the 25 seconds a renamed button costs

strix never finished one run, and our hand-run probes produced their own false positives

dsh free models: 7 of 13 answer, and the two in the headline live on another host

herdr: the work survives the client, and a restart brings back only the layout

OpenShell: 22 of 32 escapes refused, and the path its own docs say a policy cannot expose

magpie: 1.26 ms per request through the local gateway, and 1 of 4 agent configs comes back byte-identical

ponytail: 66% less code on a real repo, and the one task where it wrote nothing at all

DBX as an agent database layer: we measured the query path and the write gates

Is DBX really a 25 MB database client? We measured it on a Mac

Scraping and reach

firecrawl vs Scrapling vs Crawl4AI on the same ten pages

caveman vs rtk vs Scrapling on the same targets: what each one shrinks

Agent Reach: 2 of 5 targets read with no key, and the Reddit route its own docs rule out

Scrapling: three fetchers, 900 MB of install, and one Cloudflare page that opened

We ran Moli and headless Chrome side by side on a Mac, here is what we measured

Media and rendering

yoinks: 6 of 8 URLs, +48.7 s per file, and a picker that says mp4 when the file is AVI

hyperframes: the same 27-second clip came out as two different files in four renders, and neither of them took 4 seconds

fframes: 27.4 seconds of video in 4.3 seconds, and the encoder fallback that fooled us first