March 27, 2026 • 11 mins
My recap of virtual presentations, live streams, and workshop support — sharing wins, lessons from a humbling live coding session, and a fascinating article on solving long-running LLM memory problems.
Highlights:
- Delivered a virtual meetup for the San Francisco ACM on building knowledge graphs with the Neo4j GraphRAG Python package (code repository)
- Helped as a TA during a Road to Nodes AI workshop covering MCP server integrations with Neo4j.
- Attempted a live stream refactoring Postgres to Neo4j using OGM — ran into challenges that revealed documentation gaps and learning opportunities
- Progress on the AI Java book with a productive working session
- Shared a blog post by James Dunham: "Long Running LLM Conversations Need Working Memory, Not Just More Context" — which mirrors issues encountered in a prior RPG project where LLMs lost story continuity over time
- Upcoming: Road to Nodes AI workshop on long-term memory & agentic workflows (free, virtual)
- Upcoming: NODES AI virtual conference — April 15th (free)
Listen
Watch
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Jennifer Reif (00:05):
You are listening to the Breaktime Tech Talks podcast, a bite-sized
tech podcast for busy developers wherewe'll briefly cover technical topics, new
snippets, and more in short time blocks.
I'm your host, Jennifer Reif, anavid developer and problem solver
with special interest in data,learning, and all things technology.
Sometimes you win some andsometimes you lose some.
(00:27):
I had several successful ventures andsome not so successful ventures this week.
I had a virtual presentation, alive stream, and answered some
questions for a virtual workshop.
Plus I got a few other small thingsknocked out this week along the way.
I'll dive into what worked and whatdidn't in today's episode and bring
you along for the somewhat bumpy ride.
To start us off, early in theweek, I had a virtual meetup
(00:49):
for the San Francisco ACM.
This was tons of fun, hadlots of really good questions.
Because it was Pacific time, it was alittle bit later in the evening for me.
That actually workedout really, really well.
The house was quiet, Iwas up and ready to go.
I have a GitHub repository that Iwill link in the show notes, and
the session covered the Pythonwith Neo4j GraphRAG Python package.
(01:12):
And we dealt with building aknowledge graph through that package.
So there's some knowledge graphconstruction pieces in the
Neo4j GraphRAG Python package.
And then there's alsosome retrieval pieces.
And then of course you canintegrate with agents as well.
I covered a little bit ofall of these different topics
in the session that evening.
And I actually ended up refactoring afew things that day to improve some of
(01:35):
the results and smooth some things out.
I was going through everything andI'd reviewed everything at the end of
last week and just really felt that itwasn't as high quality as what I really
wanted it to be, and it just didn'tflow as smoothly as what I wanted.
So I went back and fixed a few things.
Adjusted the model, re-importedsome things, and the changes ended
up working much, much better.
(01:55):
I thought it presented a muchbetter result on the other side,
so I was really happy with it.
There are still some further refinementsI would like to make, so hopefully I
will get another round to work on itand make even more improvements on this.
But I loved having the event there.
I had some colleagues in the room,which was, really helpful to kind
of coordinate some things and justinteract with people in their region.
(02:18):
And then I came in and did the virtualpresentation, which was a ton of fun.
Again, lots of good questions throughoutthat kind of broke things up a little bit.
There were a range of beginnerto intermediate to advanced
questions, which was really nice andeveryone took along for the ride.
So this was a ton of fun.
It was definitely somethingI would love to do.
I'll link everything in the shownotes for the GitHub repository.
Feel free to check that out.
(02:40):
This week I also participated as aTA, so answering questions and helping
out in the chat for a Road to NODES AIsession, which is one of our workshops.
Every year, we have a nodes virtualconference that happens in the fall.
This year we have opened upa spring version of it that
is focused all around AI.
And it's called Nodes ai.
(03:01):
That is coming April 15th.
Again, just like our regular nodesevent in the fall, NODES AI is free
and virtual, and it's not quite aslong as our traditional nodes event.
That event runs 24 hours.
The nodes AI rendition runs I thinkabout seven to eight hours-ish.
But again, still free, still virtual.
Lots of great content around AI.
(03:22):
And to lead up to that event, wealways have virtual workshops.
So same thing is gonna happenwith our NODES AI event as well.
We have a road to NODES AI, whichis several virtual workshops that
occur in the weeks coming up to that.
We just had one this week, and I helpedout in the chat answering some questions.
This was a really great session.
It was all about MCP andintegrating Neo4j with MCP servers.
(03:47):
Walked through all of the servers thatwe have available, and then presented
two different demos, showcasing someworkflows and how everything operates
together and what it was doing.
This was a really interesting session, andit was really well done by the presenters.
There were two of them, bothinternal to Neo4j, both very skilled.
And you could tell just, reallyenjoying themselves there.
There were lots of great questions andvery interested attendees, so this was a
(04:09):
lot of fun for me to sit back and listento, as well as just to pop into the chat
and answer some things along the way.
There is one more workshop comingnext week, and this one is gonna be on
long-term memory and agentic workflows.
So if you're interested in registeringfor that, again, free, entirely virtual.
Don't wait to register.
I will leave the registrationlink in the show notes for that.
(04:31):
And then of course, if you're interestedin attending the Nodes AI event, which
is coming up in about three weeks now,that is entirely free and virtual.
I will leave a signup link for that too.
Then I had a live stream actually justfollowing the, the workshop session
that didn't really go according to plan.
We were refactoring Postgresto Neo4j, and we probably just
(04:53):
attempted too much all at once.
We swapped out some dependencies.
That worked fine.
We tried using OGM, which isthe Neo4j object graph mapper.
But I think that actually got us ina little bit more trouble because
there were several things that couldhave been a bit more direct if we'd
just gone with the driver where OGMabstracted a few things a bit more.
(05:14):
This was a really humbling experience.
That showed me some gaps in myunderstanding and gaps that I
am currently trying to fill.
So the live stream wasn't supersuccessful, but I'm working through
fixing what we missed, and I am reallyexcited to learn and dig in on some of
the things that I just felt like I wasn'tsure about or wasn't sure how to, get the
problem solved and where to look for it.
(05:35):
I have also noticed there's some gaps insome of the documentation, which I've,
i've mentioned before where there's nottoo many examples out there of saving
graph data to a Neo4j database, and alsocombining all these different things
where you're trying to save data to agraph database, and you're also trying to
create embeddings, and you're trying todo retrieval all in the same application.
(05:56):
I'm hoping to fill those gapshere shortly, but this was just
something that was surfaced evenmore through this experience.
I hopefully will have better updates onthis soon, and of course, share what I've
learned and how I've filled in those gaps.
I did also get a chance to workon the AI Java book this week,
and it is going really well.
I had a really good workingsession earlier in the week.
I hope to continue that over thenext, several days and weeks as well,
(06:20):
and have some more updates for that.
There was a piece of content that I got tothis week, which I'm super excited about.
I caught a few other things aswell, but I'm in the process
of working through those.
But the one that I will share thisweek, I was actually shared to me by a
colleague, the person that I did the jointpresentation with at DevNexus, Erin, sent
over this blog post called Long RunningLLM Conversations Need Working Memory,
(06:43):
not just More Context by James Dunham.
This actually hit the nail righton the head of the problems we were
running into with the joint sessionthat we did with the, the project.
This is the exact problem we saw afew months ago where we were working
on a role playing game and trying touse an LLM as the kind of game master
(07:05):
and decider and keeper of context andrules and all sorts of other things.
And the author of this article says theywere seeing the same problem where the
large language model was losing the story,contradicting some of the facts, and
confusing the character details and names.
Again, all of these things we wereseeing in our application as well.
Where RAG and memory would help.
(07:25):
But that still happened overtime, so it deteriorated as the
conversation and the game continued.
We again saw this as well.
It would do really wellupfront, and then it would just
deteriorate as the game went on.
This does make sense, andthe author's approach is very
interesting to solving this problem.
The author used Embabel as the Java AIagentic framework, and incorporated the
(07:49):
DICE piece that is their knowledge graphimplementation, for working with, memory
and knowledge graphs inside Embabel.
And there's a couple other things thatthe author mentions he used as well.
And this is something, the, theissues that the author was seeing
and that we saw as well Is somethingthat humans do automatically where
we extract key bits of facts and dataand to keep that story thread going.
(08:14):
So we automatically prioritizethings that are most important.
For instance, if somebody gets up andgives you a a 45 minute, 50 minute
presentation, you will automaticallypick out key things that you feel
are the key bullet points, the highpoints, the highlights of that.
And most people probably comeup with the same highlights.
There might be a few nuances, right?
(08:37):
But being able to keep that threadand that storyline is something
that humans do already withouteven consciously thinking about it.
And this is something thatlarge language models have not
mastered, at least not yet.
Large language models can do some ofthese pieces, but again, it erodes
over time and it can be manipulated.
Because again, large language modelsare designed to be helpful and
(08:57):
accommodating in a lot of cases.
That means that sometimes they defaultto the human, and if the human tries to
basically hack their perceptions and hackthe facts that they, that they realize
at that time or that they see as facts atthat time, then the human can come in and
manipulate those, or some other system,could potentially come in and manipulate
(09:19):
those, and the large language model willfold and cave to those, at certain points.
And this is specifically what the authorwas pointing out, is that that will erode
over time and the way sometimes that theconversation flows could erode those facts
over time or change concepts of the story.
And so the challenge then is figuringout which pieces of the story should
(09:43):
not erode over time, or shouldbe maybe slowly deteriorated and
changed over time, but shouldn'thave drastic changes all of a sudden.
So the author approached thisproblem with a ranking and evaluation
system that did improve the longrunning conversation continuity.
Again, the author saysthat it's not flawless yet.
(10:03):
It still hasn't been completelysolved, but it is doing better.
And this ranking system helped to maintainthat continuity, put up some additional
boundaries, and let the large languagemodel figure out, okay, now I have a brand
new fact that contradicts what I know.
How do I evaluate that?
Is it still.
Is that a valid change to thefacts that I know or, is it not?
(10:24):
Do I need to refute thator push back on that?
I found this really interestingand something that we'll see
how it develops over time.
But I thought this was something thatI would love to apply to some feature
projects or at least explore more.
This week, I had some successesand some challenges, but I will
spend some time working to fillsome of the gaps that I found.
I still really enjoyed the learning duringthe live stream, and I enjoyed getting to
(10:47):
answer questions and engage with attendeesduring the Road to NODES AI workshop.
As always, thanks forlistening and happy coding.
You are listening to the Breaktime Tech Talks podcast, a bite-sized
tech podcast for busy developers wherewe'll briefly cover technical topics, new
snippets, and more in short time blocks.
I'm your host, Jennifer Reif, anavid developer and problem solver
with special interest in data,learning, and all things technology.
Sometimes you win some andsometimes you lose some.
(00:27):
I had several successful ventures andsome not so successful ventures this week.
I had a virtual presentation, alive stream, and answered some
questions for a virtual workshop.
Plus I got a few other small thingsknocked out this week along the way.
I'll dive into what worked and whatdidn't in today's episode and bring
you along for the somewhat bumpy ride.
To start us off, early in theweek, I had a virtual meetup
(00:49):
for the San Francisco ACM.
This was tons of fun, hadlots of really good questions.
Because it was Pacific time, it was alittle bit later in the evening for me.
That actually workedout really, really well.
The house was quiet, Iwas up and ready to go.
I have a GitHub repository that Iwill link in the show notes, and
the session covered the Pythonwith Neo4j GraphRAG Python package.
(01:12):
And we dealt with building aknowledge graph through that package.
So there's some knowledge graphconstruction pieces in the
Neo4j GraphRAG Python package.
And then there's alsosome retrieval pieces.
And then of course you canintegrate with agents as well.
I covered a little bit ofall of these different topics
in the session that evening.
And I actually ended up refactoring afew things that day to improve some of
(01:35):
the results and smooth some things out.
I was going through everything andI'd reviewed everything at the end of
last week and just really felt that itwasn't as high quality as what I really
wanted it to be, and it just didn'tflow as smoothly as what I wanted.
So I went back and fixed a few things.
Adjusted the model, re-importedsome things, and the changes ended
up working much, much better.
(01:55):
I thought it presented a muchbetter result on the other side,
so I was really happy with it.
There are still some further refinementsI would like to make, so hopefully I
will get another round to work on itand make even more improvements on this.
But I loved having the event there.
I had some colleagues in the room,which was, really helpful to kind
of coordinate some things and justinteract with people in their region.
(02:18):
And then I came in and did the virtualpresentation, which was a ton of fun.
Again, lots of good questions throughoutthat kind of broke things up a little bit.
There were a range of beginnerto intermediate to advanced
questions, which was really nice andeveryone took along for the ride.
So this was a ton of fun.
It was definitely somethingI would love to do.
I'll link everything in the shownotes for the GitHub repository.
Feel free to check that out.
(02:40):
This week I also participated as aTA, so answering questions and helping
out in the chat for a Road to NODES AIsession, which is one of our workshops.
Every year, we have a nodes virtualconference that happens in the fall.
This year we have opened upa spring version of it that
is focused all around AI.
And it's called Nodes ai.
(03:01):
That is coming April 15th.
Again, just like our regular nodesevent in the fall, NODES AI is free
and virtual, and it's not quite aslong as our traditional nodes event.
That event runs 24 hours.
The nodes AI rendition runs I thinkabout seven to eight hours-ish.
But again, still free, still virtual.
Lots of great content around AI.
(03:22):
And to lead up to that event, wealways have virtual workshops.
So same thing is gonna happenwith our NODES AI event as well.
We have a road to NODES AI, whichis several virtual workshops that
occur in the weeks coming up to that.
We just had one this week, and I helpedout in the chat answering some questions.
This was a really great session.
It was all about MCP andintegrating Neo4j with MCP servers.
(03:47):
Walked through all of the servers thatwe have available, and then presented
two different demos, showcasing someworkflows and how everything operates
together and what it was doing.
This was a really interesting session, andit was really well done by the presenters.
There were two of them, bothinternal to Neo4j, both very skilled.
And you could tell just, reallyenjoying themselves there.
There were lots of great questions andvery interested attendees, so this was a
(04:09):
lot of fun for me to sit back and listento, as well as just to pop into the chat
and answer some things along the way.
There is one more workshop comingnext week, and this one is gonna be on
long-term memory and agentic workflows.
So if you're interested in registeringfor that, again, free, entirely virtual.
Don't wait to register.
I will leave the registrationlink in the show notes for that.
(04:31):
And then of course, if you're interestedin attending the Nodes AI event, which
is coming up in about three weeks now,that is entirely free and virtual.
I will leave a signup link for that too.
Then I had a live stream actually justfollowing the, the workshop session
that didn't really go according to plan.
We were refactoring Postgresto Neo4j, and we probably just
(04:53):
attempted too much all at once.
We swapped out some dependencies.
That worked fine.
We tried using OGM, which isthe Neo4j object graph mapper.
But I think that actually got us ina little bit more trouble because
there were several things that couldhave been a bit more direct if we'd
just gone with the driver where OGMabstracted a few things a bit more.
(05:14):
This was a really humbling experience.
That showed me some gaps in myunderstanding and gaps that I
am currently trying to fill.
So the live stream wasn't supersuccessful, but I'm working through
fixing what we missed, and I am reallyexcited to learn and dig in on some of
the things that I just felt like I wasn'tsure about or wasn't sure how to, get the
problem solved and where to look for it.
(05:35):
I have also noticed there's some gaps insome of the documentation, which I've,
i've mentioned before where there's nottoo many examples out there of saving
graph data to a Neo4j database, and alsocombining all these different things
where you're trying to save data to agraph database, and you're also trying to
create embeddings, and you're trying todo retrieval all in the same application.
(05:56):
I'm hoping to fill those gapshere shortly, but this was just
something that was surfaced evenmore through this experience.
I hopefully will have better updates onthis soon, and of course, share what I've
learned and how I've filled in those gaps.
I did also get a chance to workon the AI Java book this week,
and it is going really well.
I had a really good workingsession earlier in the week.
I hope to continue that over thenext, several days and weeks as well,
(06:20):
and have some more updates for that.
There was a piece of content that I got tothis week, which I'm super excited about.
I caught a few other things aswell, but I'm in the process
of working through those.
But the one that I will share thisweek, I was actually shared to me by a
colleague, the person that I did the jointpresentation with at DevNexus, Erin, sent
over this blog post called Long RunningLLM Conversations Need Working Memory,
(06:43):
not just More Context by James Dunham.
This actually hit the nail righton the head of the problems we were
running into with the joint sessionthat we did with the, the project.
This is the exact problem we saw afew months ago where we were working
on a role playing game and trying touse an LLM as the kind of game master
(07:05):
and decider and keeper of context andrules and all sorts of other things.
And the author of this article says theywere seeing the same problem where the
large language model was losing the story,contradicting some of the facts, and
confusing the character details and names.
Again, all of these things we wereseeing in our application as well.
Where RAG and memory would help.
(07:25):
But that still happened overtime, so it deteriorated as the
conversation and the game continued.
We again saw this as well.
It would do really wellupfront, and then it would just
deteriorate as the game went on.
This does make sense, andthe author's approach is very
interesting to solving this problem.
The author used Embabel as the Java AIagentic framework, and incorporated the
(07:49):
DICE piece that is their knowledge graphimplementation, for working with, memory
and knowledge graphs inside Embabel.
And there's a couple other things thatthe author mentions he used as well.
And this is something, the, theissues that the author was seeing
and that we saw as well Is somethingthat humans do automatically where
we extract key bits of facts and dataand to keep that story thread going.
(08:14):
So we automatically prioritizethings that are most important.
For instance, if somebody gets up andgives you a a 45 minute, 50 minute
presentation, you will automaticallypick out key things that you feel
are the key bullet points, the highpoints, the highlights of that.
And most people probably comeup with the same highlights.
There might be a few nuances, right?
(08:37):
But being able to keep that threadand that storyline is something
that humans do already withouteven consciously thinking about it.
And this is something thatlarge language models have not
mastered, at least not yet.
Large language models can do some ofthese pieces, but again, it erodes
over time and it can be manipulated.
Because again, large language modelsare designed to be helpful and
(08:57):
accommodating in a lot of cases.
That means that sometimes they defaultto the human, and if the human tries to
basically hack their perceptions and hackthe facts that they, that they realize
at that time or that they see as facts atthat time, then the human can come in and
manipulate those, or some other system,could potentially come in and manipulate
(09:19):
those, and the large language model willfold and cave to those, at certain points.
And this is specifically what the authorwas pointing out, is that that will erode
over time and the way sometimes that theconversation flows could erode those facts
over time or change concepts of the story.
And so the challenge then is figuringout which pieces of the story should
(09:43):
not erode over time, or shouldbe maybe slowly deteriorated and
changed over time, but shouldn'thave drastic changes all of a sudden.
So the author approached thisproblem with a ranking and evaluation
system that did improve the longrunning conversation continuity.
Again, the author saysthat it's not flawless yet.
(10:03):
It still hasn't been completelysolved, but it is doing better.
And this ranking system helped to maintainthat continuity, put up some additional
boundaries, and let the large languagemodel figure out, okay, now I have a brand
new fact that contradicts what I know.
How do I evaluate that?
Is it still.
Is that a valid change to thefacts that I know or, is it not?
(10:24):
Do I need to refute thator push back on that?
I found this really interestingand something that we'll see
how it develops over time.
But I thought this was something thatI would love to apply to some feature
projects or at least explore more.
This week, I had some successesand some challenges, but I will
spend some time working to fillsome of the gaps that I found.
I still really enjoyed the learning duringthe live stream, and I enjoyed getting to
(10:47):
answer questions and engage with attendeesduring the Road to NODES AI workshop.
As always, thanks forlistening and happy coding.
Popular Podcasts
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com
Betrayal Weekly
Betrayal Weekly is back for a new season. Every Thursday, Betrayal Weekly shares first-hand accounts of broken trust, shocking deceptions, and the trail of destruction they leave behind. Hosted by Andrea Gunning, this weekly ongoing series digs into real-life stories of betrayal and the aftermath. From stories of double lives to dark discoveries, these are cautionary tales and accounts of resilience against all odds. From the producers of the critically acclaimed Betrayal series, Betrayal Weekly drops new episodes every Thursday. If you would like to share your story, you can reach out to the Betrayal Team by emailing them at betrayalpod@gmail.com and follow us on Instagram at @betrayalpod and @glasspodcasts. Please join our Substack for additional exclusive content, curated book recommendations, and community discussions. Sign up FREE by clicking this link Beyond Betrayal Substack. Join our community dedicated to truth, resilience, and healing. Your voice matters! Be a part of our Betrayal journey on Substack.