Improved AI Memory? 🧠Full Hermes Tutorial (Mnemosyne & Hindsight) — Transcript
Full transcript
- 0:00So, to me this is pretty wild that not
- 0:01only are we able to get the agent to
- 0:03remember things, but it's able to build
- 0:04this graph based on our memory in a way
- 0:06that improves the agent on the problems
- 0:08I'm trying to solve. Have you ever been
- 0:10frustrated that your AI just forgets
- 0:12things? You explain the same context
- 0:14over and over again, wasting tokens and
- 0:17time. And if you're running multiple
- 0:18agents, it gets even worse because you
- 0:20have to share context just to keep them
- 0:22on the same page. That's because you're
- 0:23not using a real memory system, or at
- 0:26least not a very good one. Let's fix
- 0:28that. Hi, my name is Callum, also known
- 0:30as Waterlitz, and welcome to today's
- 0:31video on giving your agent true memory.
- 0:34Today I'm using Hermes agent as an
- 0:35example, but the memory tools I talk
- 0:37about apply to any different agentic AI.
- 0:39And more importantly, they work between
- 0:41any AI agents. You can set this system
- 0:43up once and then connect it to any agent
- 0:45with the same universal memory. If
- 0:47you're new to Hermes and want to get it
- 0:48set up safely, I have a full agentic AI
- 0:50playlist here. By the end of this video,
- 0:52you'll understand the different tiers of
- 0:54the memory stack and how agentic memory
- 0:55compares to a knowledge layer like
- 0:57Obsidian LLM Wiki. Why you might want a
- 0:59dedicated memory provider instead of
- 1:01using built-in memory, the different
- 1:02memory options and why I picked the two
- 1:04we're setting up today, how to install
- 1:05an unofficial lightweight memory layer,
- 1:07and how to install an official
- 1:09heavyweight memory engine. I'm
- 1:11personally very excited because memory
- 1:12is something I've been looking at for a
- 1:13long time, and I can't wait to help you
- 1:15get it set up today. Now, let's take a
- 1:17look at agentic memory.
- 1:22Before we get anything installed, I just
- 1:24want to quickly clear up some common
- 1:25confusions on agentic memory. It's not a
- 1:28layer, it's a stack. When people say
- 1:30second brain or agentic memory, they're
- 1:32actually referring to three things:
- 1:34world knowledge, built-in memory, and
- 1:36external memory. The knowledge layer
- 1:38like Obsidian vaults or an LLM Wiki is
- 1:40part of the memory stack, but it's not
- 1:41true operational memory the way we mean
- 1:43it when we say agent memory. It's world
- 1:45knowledge, the ground truth across every
- 1:47project and every agent. You can even
- 1:49limit the access your agent has to this
- 1:51knowledge layer, like I talked more
- 1:52about in connecting agentic AI to
- 1:54Obsidian here. Then we have operational
- 1:56memory, which splits into two tiers,
- 1:58built-in memory and dedicated memory
- 2:00providers. Operational memory focuses
- 2:02more on what the agent should remember
- 2:03about working with me specifically. A
- 2:05simple test is to consider if it's
- 2:07knowledge you want preserved and
- 2:08organized across everything you do,
- 2:10that's Obsidian. If it's a fact about
- 2:12how you like to work or what you're
- 2:13working on right now, that's agentic
- 2:15memory. So, putting it all together, we
- 2:17have built-in memory, which is great for
- 2:18simple single session chats, a memory
- 2:20provider when you're running complex
- 2:22ongoing projects or you just want
- 2:23noticeably better performance and agent
- 2:25recall, and your knowledge layer, like
- 2:27Obsidian, for building world knowledge
- 2:29that spans every project and every agent
- 2:31you run. I talked more about getting
- 2:32Obsidian connected to agentic AI in this
- 2:34video here. So, today I want to focus
- 2:36more specifically on the first two
- 2:37layers, on the built-in memory and the
- 2:39external memory provider. My next video
- 2:42is going to go deeper into setting up an
- 2:43Obsidian LLM wiki that complements the
- 2:45agent memory we're setting up today.
- 2:47Let's start by comparing the built-in
- 2:48memory to true memory providers. If you
- 2:50find this video helpful, please like,
- 2:52hype, and subscribe as I appreciate your
- 2:53support a lot. If you're looking for
- 2:55more ways to support me, please consider
- 2:56joining my YouTube or my Patreon
- 2:57membership, where I give my members
- 2:59access to tips, insights, and my
- 3:01thoughts on the future of agentic AI and
- 3:03knowledge management.
- 3:08So, what does built-in memory actually
- 3:09look like for Hermes? It has three
- 3:11files, memory.md, which is your
- 3:13operational memory, user.md, which is
- 3:15who you are, and soul.md, which is
- 3:18technically not memory but the agent's
- 3:19own personality. On top of that, the
- 3:21built-in memory includes session search,
- 3:23stored in a local database, so you can
- 3:25literally ask Hermes to search across
- 3:27your past sessions. As a quick tip, the
- 3:29best way to get going with an agent like
- 3:31Hermes is to tell Hermes about yourself
- 3:33right off the bat. If it understands a
- 3:35little bit more about who you are and
- 3:37how you like to work, the whole system
- 3:38gets so much smarter almost immediately.
- 3:41One question I get a lot about running
- 3:42the built-in memory is if you're using
- 3:44Hermes terminal execution back end,
- 3:46you're running the code in something
- 3:48like Docker, can Hermes still access the
- 3:50memory? And the answer is yes, it can.
- 3:52Running in Docker doesn't impact Hermes'
- 3:54ability to access its own memory and
- 3:56user files. That's done using a memory
- 3:58tool, which is the same thing that the
- 4:00memory providers use, rather than
- 4:01running through code that ends up being
- 4:03sandboxed in Docker. I talked more about
- 4:05that in my how to get your agent
- 4:07isolated and operating more safely on
- 4:09your computer. The key is that the
- 4:10built-in user and memory files get
- 4:12injected into every new session. It
- 4:14becomes part of the context of every
- 4:16chat you have with Hermes. Now, this is
- 4:18great because it means that your agent
- 4:19never forgets who you are, but it also
- 4:21means that it's using up some of the
- 4:22context and tokens for every single
- 4:24chat. You want to keep this as lean as
- 4:26possible, so you're not wasting tokens.
- 4:28And that's where having an actual memory
- 4:29provider comes in. A memory provider is
- 4:31an external layer that stores and
- 4:32retrieves facts at runtime or when the
- 4:34agent actually needs it, rather than
- 4:36inserting it into every conversation.
- 4:38This helps keep your context lean and
- 4:40your facts on point for when you
- 4:41actually need them. Let's take a deeper
- 4:43look into memory providers.
- 4:48A lot of people feel overwhelmed when
- 4:50trying to pick a memory provider, and
- 4:52that's fair. There's a lot of options
- 4:53and things to consider. It depends on
- 4:55your goals, how many agents or people
- 4:57are accessing that shared memory, what
- 4:59hardware you have available to run it,
- 5:00and whether you want to use free or paid
- 5:02tools. The good news is that most, if
- 5:04not all of them, allow you to migrate
- 5:05between each other, so you're not locked
- 5:07into one that you pick up front. You can
- 5:09test and experiment, and if you don't
- 5:11like one tool, you can always migrate
- 5:13your memories to another one. But rather
- 5:14than comparing them all in this video,
- 5:16which would take a while, I put together
- 5:17my analysis free Patreon post that I'll
- 5:20link in the top comment of the video, so
- 5:22you can take a look at the different
- 5:23options and see what works best for your
- 5:24scenario. But today, rather than going
- 5:26off of benchmarks or feature lists, I
- 5:28thought I would take a look at what the
- 5:29community has been rating the most
- 5:30highly and what many people have
- 5:32recommended to me. And they are Nemesis
- 5:34and Hindsight. Nemesis is a
- 5:35zero-dependency lightweight system and
- 5:37doesn't require an LLM to work. It runs
- 5:40fully locally, incredibly fast with
- 5:42built-in embeddings. Nemesis is great if
- 5:43you want a memory provider that doesn't
- 5:45slow your system down. Hindsight has
- 5:47more features and it ranks at or near
- 5:48the top of most memory benchmarks, but
- 5:50it either needs a cloud connection or
- 5:52your own server plus an LLM to actually
- 5:54power it. The simplest way to think of
- 5:56it is Nomasyne is a memory layer,
- 5:58Hindsight is a memory engine. So, let's
- 6:00start with the lightweight option of
- 6:01Nomasyne so you can see a memory
- 6:02provider in action before we move on to
- 6:05a little more power with Hindsight.
- 6:11So, we can think of Nomasyne as a
- 6:12universal memory layer for any AI agent
- 6:15running a SQLite backend with
- 6:17sub-millisecond zero dependencies. So,
- 6:19the purpose of showing you this one
- 6:21first is that it is very fast and it
- 6:24doesn't require you to install a lot of
- 6:26other things. So, a little bit more
- 6:27about Nomasyne is that it's biologically
- 6:30inspired by the BEAM architecture, which
- 6:32brings in different memory tiers like
- 6:33working, episodic, semantic, and a
- 6:35scratchpad where each one is optimized
- 6:38for different access patterns.
- 6:39Everything stays local, which is good
- 6:41for privacy. It has sub-100 millisecond
- 6:43queries and it has native Hermes
- 6:45integration.
- 6:49Let's take a look at the architecture
- 6:50for a moment. So, the key is that we're
- 6:52building auto context injection, which
- 6:54basically means that when you have a
- 6:56query, when you ask Hermes something,
- 6:59Nomasyne will receive that query, use a
- 7:02recall tool to go take a look at its
- 7:04existing context, and inject that
- 7:06context into the agent's prompts before
- 7:09it responds to you. As you continue
- 7:11having conversations, it's going to
- 7:12remember more, which is going to get fed
- 7:14back into Nomasyne to help it build this
- 7:16context assembly. So, this is just one
- 7:18great way to continue building out a
- 7:20self-evolving memory system for your
- 7:21agent. And like we can see, it can be
- 7:23used for Claude code, for Codex, it can
- 7:25be used for Cursor, it can be used for
- 7:27any agent that you want, and we can have
- 7:29all of them connect to the same memory
- 7:31bank if we want to. All right, so let's
- 7:33get Nomasyne set up.
- 7:37So, if we go to their docs here, we can
- 7:38see that they have Hermes agent
- 7:39integration and it gives a list of quick
- 7:41commands here, so it's pretty easy to
- 7:43install. And one thing I want to mention
- 7:45just very quickly is that if you're
- 7:46running Hermes and you have your
- 7:47terminal execution back end running in
- 7:49Docker, we can't just give Hermes the
- 7:51task of setting up Mnemosyne for us
- 7:53because Hermes would try and run its
- 7:55terminal, and it would run inside of a
- 7:57Docker container here, and that's not
- 7:59what we need because we need our memory
- 8:01system to survive outside of the
- 8:03terminal back end Docker container. If
- 8:05you're not running inside a Docker, you
- 8:06can just ask Hermes to do it, but
- 8:07basically this just means that we need
- 8:09to run it from our host device inside of
- 8:11our computer. So, I'm going to install
- 8:12it directly from terminal myself.
- 8:14Mnemosyne is not a built-in provider, so
- 8:17we need to install Mnemosyne a little
- 8:19bit more manually. But the key here is
- 8:21that we're not installing it inside of
- 8:23Hermes managed virtual environment
- 8:26because every time you run Hermes
- 8:27update, it rebuilds that virtual
- 8:29environment, and then it wipes any extra
- 8:31packages. So, we have two options that
- 8:33are listed here. We can use pipx, which
- 8:35creates a separate virtual environment,
- 8:37or we can manually create a virtual
- 8:40environment using the fallback option.
- 8:41However, neither of these options worked
- 8:43perfectly for me. So, if you install
- 8:45Mnemosyne and you run Hermes memory
- 8:48status and you get a plugin missing
- 8:50error, or when you first install it, the
- 8:52Python environments don't match error,
- 8:54you might need to add another step in
- 8:56between.
- 8:59So, the first step is going to be to
- 9:01install Mnemosyne Hermes into its own
- 9:03pipx environment, then set the config.
- 9:06So, these two steps are the same, but
- 9:07then before we restart the gateway, we
- 9:09need to link the new environment that we
- 9:11just created into the Hermes plugins
- 9:14folder, and then we need to run what's
- 9:15called the Hermes memory setup. That's
- 9:18just a very easy way that you can run it
- 9:19inside of terminal, and it will go
- 9:21through and configure Mnemosyne for you
- 9:23so that it gets recognized as a memory
- 9:24provider instead of just as a plugin.
- 9:26So, let's get started. Okay, so I'm
- 9:27going to go to the documentation. I'm
- 9:29going to copy the install. I'm going to
- 9:31go to my terminal and paste this in.
- 9:33Great, there we go. That took less than
- 9:35a minute. So, we now have Mnemosyne
- 9:36Hermes set up. Now, if we go over to
- 9:38memory and context, we can see that we
- 9:40currently have the default memory system
- 9:42enabled, and this is the built-in memory
- 9:44provider. But, we can see here, if we
- 9:45click the drop-down, we don't have
- 9:47access to Mnemosyne inside of here.
- 9:49That's because it's a plugin, it's not a
- 9:52built-in memory provider. So, we need to
- 9:54enable that for our default profile
- 9:56outside of the settings of the desktop
- 9:58app. So, let's do that for a moment. So,
- 9:59I'm going to go back to the docs here,
- 10:02copy this string, paste it in. Great,
- 10:04there we go. That just took 2 seconds.
- 10:06And basically, what just happened there
- 10:07is it modified the config.yaml file to
- 10:10use Mnemosyne as the memory provider.
- 10:13Now, we just installed Mnemosyne into
- 10:16One Bin, and we want to create a symlink
- 10:19or a link between where we just
- 10:21installed it using that separate
- 10:22environment and the actual Hermes plugin
- 10:25folder right here. So, that's what we
- 10:26can see right here. It just went through
- 10:28and verified that the provider of
- 10:30Mnemosyne is available and that it was
- 10:32able to connect Mnemosyne to Hermes. So,
- 10:34our next step is to run Hermes memory
- 10:36setup, and this is where we can see
- 10:37here, we want to go to Mnemosyne local,
- 10:40click yes. So, now that we are running
- 10:42Hermes memory setup, we're going to go
- 10:44through step-by-step and select all of
- 10:46the different configurations we want for
- 10:48Mnemosyne as Hermes memory provider. I
- 10:50just used the default option for
- 10:52everything except for being able to
- 10:54access global rather than session
- 10:56memories. But, I do recommend reading
- 10:58what it's asking, and if you're not
- 10:59sure, asking Hermes to help you
- 11:01understand the setup in a way that's
- 11:02going to work best for you. The most
- 11:04important setting, what's the default
- 11:05scope for remembering? I don't want to
- 11:07limit it per session, I want to have it
- 11:10be across all sessions. That's one of
- 11:11the benefits of running a memory system,
- 11:13so it's really important to switch this
- 11:15one to global. And there we go. So, we
- 11:17can see that this just went through and
- 11:18it saved the memory provider of
- 11:20Mnemosyne. So, now what we do is, if we
- 11:22go over to config.yaml inside of our
- 11:24.hermes folder, there we go. We can see
- 11:26that we have a whole memory system here
- 11:28with all of these different skills and
- 11:30tools set up for the Mnemosyne provider.
- 11:32So now I'm going to restart the Hermes
- 11:34Gateway. Going to type in Hermes memory
- 11:36status and we can see that the memory
- 11:38tools enabled, the provider is Nemesis,
- 11:40we have all of the configurations that
- 11:41we just went through so we can change
- 11:43this later if we want. We have the
- 11:44plugin installed and the status is
- 11:46available which is key. And now we can
- 11:48see here if I type in Hermes Nemesis
- 11:50stats, everything is currently at zero
- 11:52because we haven't actually stored
- 11:53anything yet. So let's get that started.
- 11:58I'm going to open the Hermes desktop app
- 12:00and if you go over to memory and
- 12:01context, we can see that Nemesis is now
- 12:03set up as the memory provider. Great. So
- 12:05now we have it all installed and if we
- 12:07ever want to upgrade this, we can just
- 12:08type in pipx upgrade Nemesis-Hermes.
- 12:11One more step I want to quickly mention
- 12:13here is that the Nemesis documentation
- 12:15suggests disabling legacy memory. And
- 12:17basically what that means is just this
- 12:19persistent memory and user profile we
- 12:21have set. The main reason is to avoid
- 12:23duplication and token waste. So I would
- 12:25suggest disabling it to start and then
- 12:28testing it out and if you notice a
- 12:29problem, you can always re-enable it.
- 12:31But don't disable it using a command
- 12:33like this. Make sure you just update it
- 12:35inside of the config.yaml file where we
- 12:38can see memory enabled false, user
- 12:39profile enabled false, which is the same
- 12:42thing as toggling these buttons off
- 12:44right here.
- 12:47Okay, so let's quickly test it out. If I
- 12:49ask Hermes a question, it shouldn't know
- 12:51anything about me. So previously I had
- 12:53told Hermes who I was, but now why don't
- 12:55we try it again since we have Nemesis
- 12:57set up. So we can see here it's going
- 12:58and it's actually enabling the Nemesis
- 13:00recall skills and the canonical skill,
- 13:03but it doesn't have anything. So I'm
- 13:05going to give it a little bit about me
- 13:06so that it can remember it. Just as a
- 13:07quick tip, this is always the first
- 13:09thing I would do when you're setting up
- 13:10a new memory system is give Hermes or
- 13:12your agent some context on who you are
- 13:14and what you like. That way it starts to
- 13:16build a profile around you and the chat
- 13:19will start to feel more personalized and
- 13:21effective for what you're trying to
- 13:22accomplish. So my name is Callum, I go
- 13:24by Wander Roots, I run a YouTube
- 13:26channel, I'm a patent agent,
- 13:27intellectual property lawyer. I studied
- 13:29mechanical materials science
- 13:30engineering, and this is my preference
- 13:32for how I like to respond. So, we can
- 13:35see here, now Hermes is going through
- 13:37and is running the Nemesis tool. It's
- 13:39running remember. It has professional
- 13:41background, alias, and name. And all of
- 13:44these things are being stored into the
- 13:47database that Nemesis has set up. So,
- 13:50now for example, if I open a new chat
- 13:51and I say who am I, we should see Hermes
- 13:54recall from Nemesis. Here we go. It
- 13:56pulled it in. That's awesome. That means
- 13:57that it is working perfectly. And now if
- 13:59we go back to terminal and run Hermes
- 14:01Nemesis stats, we can see here there's a
- 14:02total of four pieces of working memory
- 14:05unconsolidated. So, this is where it's
- 14:07going to go through and over time it's
- 14:08going to shift working memory into
- 14:10episodic memory and then deeper into it.
- 14:12And this is bringing in what's called
- 14:14the beam architecture. So, it's going to
- 14:16observe, it's going to take in
- 14:17conversation and tool results, and it's
- 14:19going to remember all of these things
- 14:20into its working memory. Then over time
- 14:22it's going to consolidate that into its
- 14:24long-term experiences or its episodic
- 14:26memories. Every time we ask the agent
- 14:28something, every time there's an input,
- 14:30it's going to go through and pull from
- 14:31both of those and then recall it. So,
- 14:33there's more details on working memory
- 14:36but this is a pretty cool system
- 14:37especially considering that it's all
- 14:39happening locally inside that SQLite
- 14:41database. And what's cool too is it's
- 14:43also bringing in structured knowledge
- 14:45stored as subject predicate object
- 14:47triples which is the basis of a
- 14:48knowledge graph. And if you want to
- 14:50learn more about that, I do have a video
- 14:51on knowledge graphs and how we can use
- 14:53it for improving agent systems. But I
- 14:55highly recommend going through and just
- 14:56exploring the beam overview so you
- 14:58understand a little bit more how this is
- 15:00all operating. This is pretty cool.
- 15:05Remember, if we go here, we had to
- 15:06manually add Nemesis to this list. But
- 15:09there's all these built-in ones that
- 15:11perhaps are a little more powerful
- 15:12depending on what you're looking for.
- 15:14And I'm going to work on installing
- 15:15hindsight today. So, I personally like
- 15:18the fact that Nemesis can run locally,
- 15:21it runs without me needing to create
- 15:24anything extra. It doesn't have any
- 15:25dependencies and it's very fast. So, I'm
- 15:27going to leave this as my default right
- 15:28now, but I'm going to create a new
- 15:29profile and I'm going to call it
- 15:31hindsight and I'm going to use this to
- 15:33set up a separate configuration here.
- 15:35Okay, so I'm now on the hindsight
- 15:36profile and if we go over to memory and
- 15:38context again, we can see here that I
- 15:40have persistent memory and user profile
- 15:42still turned on and the built-in memory
- 15:44provider is set. So, I don't have this
- 15:47one set up to connect to Namaseni. You
- 15:50can see that the hindsight profile
- 15:51doesn't even see it here. So, what we
- 15:53can do is we can go over to memory
- 15:55provider in our settings. We can go down
- 15:57to hindsight and turn this on. But, just
- 15:59before I do that, let's take a look at
- 16:00the hindsight website itself.
- 16:06So, hindsight is a native memory
- 16:08provider in the Hermes agent, which
- 16:10means it's built-in. And hindsight is a
- 16:12little bit different because it operates
- 16:14by having the agent actually connect to
- 16:16a separate server. So, it's not
- 16:18something that's running within Hermes
- 16:19itself, but what's cool is it has a
- 16:21whole bunch of different ways of
- 16:23operating. It builds its own mental
- 16:25models. It pulls in observations. It has
- 16:28standard memories. There's a lot you can
- 16:29do here and it can even reflect and then
- 16:32it's going to go through and extract
- 16:34more aware responses using a deeper
- 16:37sense of reasoning. So, when you run
- 16:39reflect, it runs an agentic loop that
- 16:41automatically searches the memory and
- 16:43then applies that to shape its own
- 16:44reasoning style and then produces a
- 16:47final answer grounded in what it's
- 16:48found. So, rather than just returning
- 16:50raw facts, reflect is a more synthesized
- 16:53response, which is why it also requires
- 16:55an LLM. So, there's a little bit more
- 16:57complexity here. I highly recommend
- 16:59exploring the hindsight documents if you
- 17:01want to learn a little bit more about
- 17:02it, but the key here is that this
- 17:04requires actually running a separate
- 17:06server or connecting to a separate
- 17:07server. So, if we go back to our memory
- 17:09provider here and select hindsight, we
- 17:12can see that a mode appears.
- 17:16So, this mode we have the option to
- 17:18switch between cloud and local external.
- 17:20And to understand that a little bit
- 17:22more, we can go specifically to the
- 17:23Hindsight docs on setting up for Hermes
- 17:25desktop specifically. And if you want to
- 17:27do it for the command line, there's also
- 17:29a separate way to get it going using the
- 17:30Hermes memory setup like you saw we did
- 17:33for Namacini. So basically we just need
- 17:34to select Hindsight, set what mode we
- 17:36want. And here we have cloud which
- 17:39requires an API key. That's where we can
- 17:41connect to the Hindsight cloud. But this
- 17:43does have a fee to it. It has you go so
- 17:47you can start for free and then
- 17:48depending on how often you're using it,
- 17:50depending on how big your system gets,
- 17:52you are going to have to start paying
- 17:53for it. You could just run the system
- 17:55yourself for free on your own computer
- 17:57if you have the hardware to do so. But
- 17:59if you don't, it might be worth taking a
- 18:00look at Hindsight cloud to see how much
- 18:02it costs. I personally haven't tried
- 18:03this yet. And today we're going to set
- 18:05up a local system. So take a look at
- 18:07this comparison here on comparing local
- 18:10to cloud LLMs for running your agent
- 18:11memory. So this is the LLM that's going
- 18:14to power the actual system as opposed to
- 18:16connecting to Hindsight cloud which also
- 18:18deals with storage. You can just connect
- 18:20to an open AI Anthropic or Gemini API,
- 18:22but then you're going to pay per token
- 18:24for every time you retain it. So you can
- 18:26run free locally on your own computer
- 18:28using something like Ollama which I have
- 18:30a video on running Ollama for agentic AI
- 18:32if you're interested in seeing how that
- 18:33works. One thing to keep in mind is that
- 18:35not all local models support tool
- 18:37calling. So you have to make sure that
- 18:38the one you're using does. I'm going to
- 18:40use GPT-OSS 20 billion parameters which
- 18:43is 13 gigs. So make sure you have that
- 18:45model installed already. And also make
- 18:47sure that you have the Ollama app
- 18:49running in the background or serve it
- 18:51yourself from terminal.
- 18:55So our next step now that we have the
- 18:57local model downloaded is that we need
- 18:58to connect Hindsight to that local model
- 19:02and get the server running in the
- 19:04background. So I'm going to do this
- 19:05inside of a docker container because
- 19:07this just improves the safety. Right now
- 19:09I already have my Hermes terminal back
- 19:10end is connecting to a docker container,
- 19:12so we're going to add another Docker
- 19:14container here that runs our memory
- 19:16system in hindsight directly. We can
- 19:18paste in this string, which I'll include
- 19:19in the free article linked in the top
- 19:21comment, but basically we want to run
- 19:23the Docker container with the name of
- 19:25hindsight, and we want it to even if the
- 19:26computer shuts down and starts up again,
- 19:28we want it to restart on its own. We're
- 19:30giving it two different ports, one for
- 19:32running the actual server, and then one
- 19:33for us to be able to look at the
- 19:34dashboard. And then these are all just
- 19:37environment variables that we are
- 19:39configuring so that it can connect to
- 19:41specifically the model that I have just
- 19:43selected here, which is the GPT-OSS 20
- 19:45billion. Okay, there we go, and it was
- 19:47able to download hindsight for us. And
- 19:49now if we go over to Docker Desktop,
- 19:51which by the way this has to be running,
- 19:52I have a whole video that goes deeper
- 19:54into how to get your Hermes agent and
- 19:56now something like hindsight
- 19:57containerized inside of Docker, but we
- 19:58can go here and we can see that we have
- 20:00our server is now running. So if we go
- 20:02over to dashboard here, we can select
- 20:04memory bank, we can create a new one, we
- 20:06can get everything going. But this just
- 20:08shows us that we have the server running
- 20:10and it's available for us to connect to.
- 20:12Okay, so our next step is to connect
- 20:14this running server to Hermes. Let's do
- 20:16that for a moment.
- 20:20So remember I'm in my hindsight profile,
- 20:22I'm going over to memory and context. We
- 20:24have hindsight settings here since we've
- 20:26selected the drop-down, and I'm going to
- 20:27change this to local external. Because
- 20:29of that we can leave the API key blank.
- 20:31I'm going to instead of having this
- 20:32connect to hindsight cloud, I'm going to
- 20:34put in the local host server that we
- 20:36just set up. I'm going to give it the
- 20:37name of Hermes, and I'm going to keep
- 20:39this recall budget as mid. So this isn't
- 20:42doing anything to control the AI model,
- 20:44this is just telling Hermes where to
- 20:45find the hindsight program that we just
- 20:48launched. So now just to test this, I'm
- 20:49going to go back and I'm going to
- 20:51actually turn off persistent memory.
- 20:52This is something that you can run in
- 20:54the background with hindsight. I'm just
- 20:56going to turn it off for a second so we
- 20:57can test hindsight. And this is
- 20:58different than Namaste, which recommends
- 21:01disabling this. So I'm going to say who
- 21:02am I? There we go. So we can see
- 21:04hindsight recall just triggered, but it
- 21:06doesn't know who I am. So, I'm giving it
- 21:08the same prompt as before to see if it
- 21:10can remember this. We can see it's
- 21:12planning durable memory. It's trying to
- 21:14update it, but again, in memories I
- 21:16turned that off. So, now it's running
- 21:18hindsight retain, and this can take a
- 21:20little bit because I'm running a local
- 21:21model. But, if we go take a look at
- 21:23Docker for a second, we can go to my
- 21:24hindsight container. We can see here
- 21:26that it's running triggering the memory
- 21:28engine. So, if the model that I selected
- 21:30has gone to sleep, it can take a little
- 21:31bit. But, we can see here it's working
- 21:33to store in chunks. Okay, there we go.
- 21:35It was able to use hindsight retain to
- 21:38store the user profile and preferences.
- 21:40Cool. So, now it's saying too, the most
- 21:41useful optional context would be what
- 21:43are the answers to all of these things.
- 21:45But, I'm just going to start a new
- 21:46session and say, "Who am I?" and see if
- 21:48it's able to pull that information from
- 21:50hindsight. There we go. So, rather than
- 21:51retain, it's now running hindsight
- 21:53recall, and it pulls up all the details
- 21:55that we had before, which is great. That
- 21:57means that this memory provider is
- 21:59working.
- 22:02And now, if we go back to localhost 99
- 22:05dashboard, we can select the Hermes
- 22:07memory bank that now exists here. We can
- 22:09click on it, and what's super cool is we
- 22:10have the ability to see all of the
- 22:12information here that's already been
- 22:14created as part of the hindsight system.
- 22:17So, to me this is pretty wild that not
- 22:19only are we able to get the agent to
- 22:20remember things, but it's able to build
- 22:22this graph based on our memory, and
- 22:24there's so much detail we can get into
- 22:26here. This is called constellation view.
- 22:28We can switch it to table. We can take a
- 22:30look at timeline, which would happen as
- 22:32we get more details. There's
- 22:33experiences, observations, mental
- 22:35models. There's so many things that we
- 22:38can structure here that I'm personally
- 22:39super excited to dive into.
- 22:44So, for example, I'm just going to
- 22:45answer the question that Hermes had
- 22:47given me, giving it a little more
- 22:48context on what I hope to do with my
- 22:51YouTube channel, who I like to help, why
- 22:53I'm doing this in the first place. So,
- 22:55I'm going to have hindsight now retain
- 22:57this information. Okay, so it was able
- 22:59to run there, and if I click refresh on
- 23:01this graph, we can see that there's now
- 23:03a few more pieces of information. I can
- 23:05tell Hermes to please reflect on this
- 23:06conversation.
- 23:09So, here's the result of the reflection
- 23:11that also pulled in some recall. There
- 23:13was a bit of an issue, but it seems to
- 23:14have done a pretty good job pulling in a
- 23:17deeper understanding of what it is I'm
- 23:19trying to do just based on that brief
- 23:21discussion I've had. I'm pretty
- 23:22impressed that it's pulling this in so
- 23:24quickly. The capable but overloaded
- 23:26builder people that I'm chatting with
- 23:28and I'm teaching here are curious and
- 23:30are wanting to understand rather than
- 23:32just getting answers, but I hope you can
- 23:33see that not only are these memories
- 23:35being stored here, they're being used in
- 23:37a way that improves the reflection that
- 23:39the agent has on the problems I'm trying
- 23:41to solve and what I'm interested in
- 23:43doing. And that was just after a basic
- 23:45amount of information. So, that's pretty
- 23:47cool. And again, this reflection is
- 23:48something that from what I understand at
- 23:50least at this moment, Mnemosyne doesn't
- 23:52have yet. So, that brings me to the
- 23:54question of which one should you choose?
- 23:56What should we actually care about here
- 23:59and where should we take this next?
- 24:05The author of Mnemosyne says these are
- 24:07not direct competitors. Hindsight is a
- 24:09memory engine with sophisticated natural
- 24:11language processing and multi-signal
- 24:13retrieval. Mnemosyne is a memory layer
- 24:15optimized for simplicity, speed, and
- 24:17single machine deployments. So, this is
- 24:19really the summary of it. Hindsight is
- 24:21an engine, whereas Mnemosyne is a layer.
- 24:24So, it's up to you to decide what works
- 24:25best for you. I am going to put more
- 24:27information here and all of the
- 24:28resources I talk about today into a free
- 24:30Patreon article that will go more
- 24:32in-depth on how to make this decision
- 24:34yourself, think the key here is to test
- 24:36one provider for a week and see if it
- 24:38improves your workflow. If it doesn't,
- 24:40try another one. You can always export
- 24:42from one to the other, there's no
- 24:43lock-in. And again, one thing that I
- 24:45personally am excited to explore a lot
- 24:47more is the dashboard here. I just think
- 24:49this is so cool that I'm able to go and
- 24:51see all of these elements coming
- 24:53together, this full dashboard of the
- 24:55memories that are being created from my
- 24:57agent as I just have conversations with
- 24:59it. And to me, this is also the key
- 25:02benefit of running something like
- 25:03hindsight over Nemesimy because while
- 25:06Nemesimy does have a dashboard, it's a
- 25:09community dashboard. It's not built by
- 25:10the creator of Nemesimy. So, Nemesimy is
- 25:13a smaller tool built by a smaller team
- 25:15and hindsight is a full system that's
- 25:18got a lot more going for it. And that
- 25:20brings me to the next step, which I'm
- 25:21really excited about. Not just building
- 25:23memories based on my chats with the
- 25:25agent, but also connecting this to my
- 25:26world knowledge, my library, or my
- 25:29second brain inside of my Obsidian LLM
- 25:31Wiki.
- 25:34So, now your agent has true agentic
- 25:36memory. Nemesimy gives you a fast
- 25:38lightweight memory layer, and if you're
- 25:39looking for a little more power, you can
- 25:41use the memory engine of hindsight. If
- 25:43you have any questions about what we
- 25:44talked about today or tips for how
- 25:45you're using your own workflow, please
- 25:47let me know in the comments. I'm happy
- 25:48to answer questions and I love learning
- 25:50how you're using these systems as well.
- 25:52But, there's still one piece missing,
- 25:53world knowledge. Your operational
- 25:55memory, the agentic memory we set up,
- 25:57now knows how you work. Next, I'll
- 25:58connect it to the Obsidian LLM Wiki so
- 26:00your agent isn't just remembering facts
- 26:02about you. It's also reasoning against
- 26:04everything you already know. If you
- 26:06found this video helpful, please like
- 26:07and subscribe. I really appreciate it.
- 26:09And don't forget to check out the free
- 26:10Patreon post that I've linked below with
- 26:12more information on the setup, the code,
- 26:14and the memory provider options. Thanks
- 26:16again for watching and I will see you in
- 26:17the next video.
About this transcript
This page contains the full transcript of Improved AI Memory? 🧠Full Hermes Tutorial (Mnemosyne & Hindsight) by Wanderloots, generated from the public captions YouTube serves with the video. The transcript has 5,687 words across 818 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.
What you can do with it
Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.
Free YouTube transcript tool
YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.