Videos Q0VkgCyNVUg

CrabRAG: Why Automated Assistants Need Graph Memory, Not More Tokens — Stephen Chin, Neo4j

index_state ready data_status ok

AI Engineer· published 2026-07-22· 0:20:42· en-US· indexed 2026-08-10 19:47

Open on YouTube

57 shot(s).

keyframes kept every frame deduplicated

What was stored

cues: 237; whisperx 237
chunks: 35; from 237 cues
keyframes: 42; kept of 57 captured
frames with text: 41; 1,042 lines read
chapters: 12; from the source metadata
keyframe bytes: 7.5 MB; word timings on 237 cues

Provenance

Each pipeline stage, its state and the model that produced it
stage	state	model	started	took
`fetch`	done	—	2026-08-10 03:55	1m 42s
`stt`	done	—	2026-08-10 03:57	21s
`chunk`	done	—	2026-08-10 03:57	0s
`text_embed`	done	—	2026-08-10 19:46	1s
`keyframe`	done	—	2026-08-10 03:57	2m 07s
`ocr`	done	—	2026-08-10 03:59	20s
`frame_embed`	done	—	2026-08-10 19:47	7s

Frames, and what the machine read

0:02 #0 done2 line(s)

shot 0·sharpness 452.4
1. AlEngineer0.96
2. World's Fair1.00
0:03 #1 done2 line(s)

shot 1·sharpness 656.2
1. AlEngineer0.95
2. World's Fair0.99
0:10 #2 done24 line(s)

shot 2·sharpness 2733.9
1. LAB & PLATINUM SPONSORS0.99
2. Amazon AGI Lab0.98
3. ANTHROP\C1.00
4. Google DeepMind1.00
5. MINIMAX0.92
6. OpenAI0.92
7. Akamai1.00
8. arize1.00
9. aws1.00
10. Braintrust bright data0.99
11. B1.00
12. Browserbase1.00
13. docker1.00
14. :neo4j0.93
15. ORACLE1.00
16. PayPal1.00
17. qodo1.00
18. reducto1.00
19. Sonar1.00
20. Makers of1.00
21. togetherai1.00
22. Unblocked1.00
23. WorkOS1.00
24. SonarQube1.00
0:15 #3 done2 line(s)

shot 3·sharpness 202.9
1. AlEngineer0.99
2. World's Fair1.00
0:45 #4 done13 line(s)

shot 4·sharpness 2380.9
1. AlEngineer0.98
2. World'sFair1.00
3. neo4j0.98
4. Graph Intelligence1.00
5. Platform1.00
6. CrabRAG1.00
7. PRESENTED BY1.00
8. Why your assistant needs1.00
9. Microsoft1.00
10. graph memory, not more tokens.1.00
11. Stephen Chin1.00
12. VP of Developer Relations, Neo4j· @steveonjava0.99
13. Engineering the future of Al1.00
1:25 #5 done17 line(s)

shot 5·sharpness 2056.3
1. AlEngineer0.97
2. THE FORGETFUL CRAB·10.98
3. World'sFair1.00
4. Every session, he1.00
5. wakes up an amnesiac.0.97
6. PRESENTED BY1.00
7. The only reason Clawd knows his own name is0.99
8. Microsoft1.00
9. the file on the sink: SOUL. md0.96
10. No hiddlen state. An OpenClaw agent only0.99
11. remembers what's saved to disk — and1.00
12. re-reads it from zero each time.0.99
13. SOUL.md1.00
14. Name: clawd1.00
15. WorkfsFar0.56
16. TRACK 5• JULY 2, 20260.96
17. Graphs1.00
1:33 #6 skipped

shot 6·duplicate of #5
1:56 #7 done12 line(s)

shot 7·sharpness 1833.4
1. AlEngineer0.97
2. World'sFair1.00
3. THE FORGETFUL CRAB· 20.94
4. A drawer full of tools.0.99
5. Still can't eat soup.1.00
6. Screwdriver? Wrench? Clawd has every tool an0.99
7. agent could want — and reaches for the wrong0.98
8. one anyway1.00
9. The problem was never tool access. It's knowing0.98
10. which tool this moment calls for.1.00
11. TRACK 5• JULY 2, 20260.95
12. Graphs1.00
2:40 #8 done16 line(s)

shot 8·sharpness 1862.1
1. AlEngineer0.97
2. World'sFair1.00
3. HOW AGENTS WORK1.00
4. The loop is easy.0.99
5. The memory is the hard part.0.99
6. PROMPT1.00
7. THINK0.99
8. CALL A TOOL0.99
9. OBSERVE1.00
10. C0.58
11. What survives between turns? Between sessions? This is the part0.99
12. MEMORY?1.00
13. everyone hand-waves — and the part that makes assistants flaky.0.99
14. Wokiar0.52
15. TRACK 5· JULY 2,20260.95
16. Graphs1.00
2:46 #9 skipped

shot 9·duplicate of #8
3:31 #10 done12 line(s)

shot 10·sharpness 1738.1
1. AlEngineer0.98
2. THE CATCH0.99
3. World'sFair1.00
4. 5m1.00
5. No structured memory means Clawd1.00
6. re-reads the entire stack every session.0.99
7. OpenClaw is anti-RAG by design – and0.99
8. the token bill shows it.1.00
9. tokens. every day.1.00
10. Don't read more. Remember better.1.00
11. TRACK 5• JULY 2, 20260.95
12. Graphs1.00
3:54 #11 skipped

shot 11·duplicate of #10
4:21 #12 done15 line(s)

shot 12·sharpness 1847.9
1. AlEngineer0.98
2. World'sFair1.00
3. HERMES·THE SUCCESSOR1.00
4. Hermes writes its own skills.1.00
5. HERMES1.00
6. Same SOUL.md identity.Acurated,capped MEMORY.md.1.00
7. AGENT1.00
8. Plus a learning loop : do the work, notice it worked, save the recipe.0.99
9. DO THE TASK0.95
10. 3-4 WINS0.99
11. SKILL.md1.00
12. reused automatically next time0.99
13. It remembers methods, not just facts — procedural memory.0.99
14. TRACK 5· JULY 2,20260.96
15. Graphs1.00
4:39 #13 skipped

shot 13·duplicate of #12
5:12 #14 done12 line(s)

shot 14·sharpness 1766.0
1. AlEngineer0.98
2. World'sFair1.00
3. SKILLS BREAK· 2· HARD TO PICK0.95
4. He meant to jog.1.00
5. He picked waterskiing.1.00
6. Hundreds of skills in the folder. Hermes' hub alone ships ~700. Choosing1.00
7. the right one for this moment is its own hard problem.0.98
8. Wrong skill — confidently executed.0.98
9. 131.00
10. Wordiae0.67
11. TRACK 5· JULY 2, 20260.96
12. Graphs1.00
5:37 #15 done16 line(s)

shot 15·sharpness 1830.9
1. AlEnaineer0.97
2. SKILLS BREAK· 2· HARD TO PICK0.93
3. World'e Cair0.94
4. Cracks the shell.1.00
5. need1.00
6. more shells...1.00
7. Wanders off.1.00
8. PRESENTED BY1.00
9. Hehascrack_shell andeat_meat.0.98
10. Microsoft1.00
11. He just can't connect them.0.97
12. Skills don't compose . Real work is a chain of steps — and the1.00
13. meat is inside the shell he just cracked.1.00
14. (Hold that thought)0.99
15. TRACK 5· JULY 2, 20260.95
16. Graphs1.00
5:50 #16 done24 line(s)

shot 16·sharpness 2032.7
1. AlEngineer0.99
2. World's Fair0.98
3. A THIRD ASSISTANT0.99
4. Great at shrimp.0.98
5. Defeated by a clam.0.98
6. PRESENTED BY1.00
7. GooseX0.98
8. Everything is an MCP extension.1.00
9. Microsoft1.00
10. 70+1.00
11. MCP extensions —0.98
12. browsers, databases, docs,0.98
13. An open-source Al agent that runs on your machine0.98
14. tools1.00
15. – desktop app, CLl, and API. Built in Rust, works0.97
16. 100%1.00
17. with any LLM.1.00
18. local — your machine, your0.98
19. keys, your files.1.00
20. An Agentic Al Foundation (AAIF) project· Linux Foundation0.99
21. Memory is just another1.00
22. MCP server0.99
23. TRACK 5· JULY 2,20260.95
24. Graphs1.00
6:20 #17 done32 line(s)

shot 17·sharpness 2392.3
1. AlEngineer0.98
2. World's Fair0.99
3. GOOSE MEMORY1.00
4. Memory is just1.00
5. anotherMCP server.0.97
6. PRESENTED BY0.99
7. remember_memory()1.00
8. retrieve_memories()1.00
9. Memories are plain files on disk1.00
10. Microsoft1.00
11. remove specific memory()0.99
12. remove_memory_category()1.00
13. ~/.config/goose/memory/ global/0.99
14. preferences.md L local/ project-x.md0.98
15. Driven by trigger words – say0.96
16. remember...0.99
17. and it1.00
18. writes a tagged entry; say1.00
19. forget...1.00
20. and it drops1.00
21. # development #rust #tooling - prefers0.99
22. one. Stored as plain files, scoped local or global ,0.98
23. ripgrep over grep0.99
24. sorted into categories and #tags.0.99
25. local ·this project0.99
26. global·everywhere1.00
27. The same catch.1.00
28. Goose loads all saved memories into every prompt, every session. More memory means1.00
29. more tokens on every call — and still no relationships.0.99
30. Workdi Fuir0.68
31. TRACK 5· JULY 2,20260.96
32. Graphs1.00
6:34 #18 skipped

shot 18·duplicate of #17
7:18 #19 done11 line(s)

shot 19·sharpness 2413.2
1. AlEngineer0.98
2. World'sFair1.00
3. MCP MEMORY·2·TOKEN BLOAT0.98
4. Remembers everything.1.00
5. Can't take off.1.00
6. Every memory, every prompt, every time. Load the0.99
7. whole pile on each call and eventually the bird can't fly.1.00
8. Flat memory grows — it never gets lighter.0.97
9. Wordiar0.53
10. TRACK 5• JULY 2, 20260.94
11. Graphs1.00
7:27 #20 skipped

shot 20·duplicate of #19
8:06 #21 done11 line(s)

shot 21·sharpness 1347.7
1. AlEngineer0.98
2. World'sFair1.00
3. HOW SIMILARITY SEARCH WORKS1.00
4. Find the nearest points1.00
5. in space.1.00
6. Every note becomes a vector. Your query1.00
7. becomes a vector. Return the closest ones.0.98
8. Excellent for "what looks similar?"0.99
9. query → 3 nearest neighbors0.96
10. TRACK 5· JULY 2, 20260.96
11. Graphs1.00
8:34 #22 done10 line(s)

shot 22·sharpness 1697.6
1. AlEngineer0.98
2. World'sFair1.00
3. THE CATCH WITH VECTORS0.98
4. Similar ≠ related.0.97
5. Nearest ≠right.0.98
6. Embeddings know when two things look alike.0.99
7. They don't know that one caused, belongs to,0.99
8. or connects to the other0.98
9. TRACK 5· JULY 2, 20260.93
10. Graphs0.92
8:45 #23 skipped

shot 23·duplicate of #22

Transcript

237 cues· 3,054 words· 16,266 chars

chunk 0 · 0:12–0:40 · 53 words · 295 chars
0:12 My name's Steven Chin.
0:14 I run the developer relations team here at Neo4j.
0:21 And I'm excited to talk to you about something we've all come to love, our crustacean friends.
0:29 So we have OpenClaw mascot, we have a bunch of other crustaceans, and we're gonna focus on one member
0:38 of the crustacean family.
chunk 1 · 0:40–1:20 · 104 words · 532 chars
0:40 I love crabs, so our little boy, Crab D. And I think in the journey to figure out how to apply agents, how to do things which are more autonomous, we're all looking for ways where we can get better results, more accurate answers, and to actually capture all of this.
1:00 But the tools kind of work against us.
1:02 So here's our friend, Crab D. He's a personal assistant.
1:08 very happy, very eager.
1:09 He wants to help us out with our lives, maybe to help us to code, to help us to manage our email, to do different things, but he's got a problem.
chunk 2 · 1:22–2:05 · 128 words · 649 chars
1:22 And our poor boy, Crabdy, has a very bad memory.
1:27 He wakes up every day and his memory file flips and now it's a new day and he forgets everything from yesterday.
1:33 Does this happen to you where you wake up and you're using OpenClaw and suddenly it's on a new set of memory files and remembers nothing that you actually did the previous day?
1:44 He's got a lot of tools at his disposal.
1:46 I mean, we love giving our agents tools, but sometimes he doesn't pick the right tool for the job.
chunk 3 · 1:53–2:33 · 114 words · 585 chars
1:53 I don't think either of these are going to help him drink his ball of soup.
1:56 So that's not the tool which he was looking to reach for.
2:03 and a little bit forgetful at times.
2:06 I don't remember everybody I meet, but I'm pretty good at faces.
2:10 If I've met you before, I recognize faces.
2:12 It's like, pleased to meet you.
2:16 Crabd is not as good at that, so very forgetful.
2:19 It's like you're reteaching it every day to do the same sort of tasks, and we want agents which are more helpful, which are able to do more for us.
chunk 4 · 2:26–3:09 · 111 words · 634 chars
2:26 Let's dig into how Crabd actually works.
2:31 It's basically a memory loop, right?
2:34 So we're prompting, we're thinking about the response, maybe calling tools, observing what happens, but the hard part is the memory.
2:43 The hard part is what you put in context, what you're recalling from, and
2:49 The way which you have memory structured in most tools is an example of how OpenClaw structures things, is you have a sol.md for your agent's memory.
chunk 5 · 2:58–3:42 · 122 words · 656 chars
2:58 You have maybe memory files.
3:00 You have different tool files.
3:01 You have daily memory files.
3:04 Now, if you look at this, there's one thing which is in common with all of these.
3:08 They're just markdown files.
3:10 So markdown files are great.
3:12 That's easy for us to read.
3:14 We can look through it.
3:15 We can quickly figure out what's not needed and compact them.
3:19 They're intentionally small for agents because you have a limited context window and also you need to keep the right things at the top of the context.
3:28 But if your whole memory is a bunch of markdown files, you're wasting a lot of tokens.
chunk 6 · 3:35–4:18 · 123 words · 631 chars
3:35 So my average agents are loading up at least 100k in tokens for each round.
3:43 They're doing a lot of skills.
3:45 They're adding a lot of things into the context constantly.
3:47 It's very repetitive because they basically load up everything in the hopes that something will be useful in the context.
3:56 At small scale, that works where you get the results you want with a high quality model.
4:00 It doesn't work at large scale.
4:02 And I'm going to show a demo of large scale where
chunk 7 · 4:06–4:51 · 123 words · 597 chars
4:06 We take OpenClaw and we let it run loose on my home lab.
4:10 So high demo risk, but a lot of fun.
4:14 And a classic digital twin scenario.
4:16 So I think we'll have a lot of fun here.

Chapters

0:00 Meet Crab D and the agent memory problem
2:34 Why markdown memory wastes tokens
4:43 Skills are just markdown too
5:49 Goose: memory as an MCP server
7:44 Vector databases and why similarity is not a relationship
9:54 Enter graphs: precise, explainable, auditable
11:38 You do not need to be a graph expert, Claude writes Cypher
12:04 The demo: a home lab digital twin, vector versus graph
13:23 Live: finding end of life software on the network
15:30 Live: finding exposed management ports
16:49 Why large scale needs graph memory
18:05 Resources: the GraphRAG book and GraphAcademy