YouTube2Text

Improved AI Memory? 🧠 Full Hermes Tutorial (Mnemosyne & Hindsight) — Transcript

by Wanderloots · 5,687 words · 818 segments · language en · Watch on YouTube

Full transcript

  1. 0:00So, to me this is pretty wild that not
  2. 0:01only are we able to get the agent to
  3. 0:03remember things, but it's able to build
  4. 0:04this graph based on our memory in a way
  5. 0:06that improves the agent on the problems
  6. 0:08I'm trying to solve. Have you ever been
  7. 0:10frustrated that your AI just forgets
  8. 0:12things? You explain the same context
  9. 0:14over and over again, wasting tokens and
  10. 0:17time. And if you're running multiple
  11. 0:18agents, it gets even worse because you
  12. 0:20have to share context just to keep them
  13. 0:22on the same page. That's because you're
  14. 0:23not using a real memory system, or at
  15. 0:26least not a very good one. Let's fix
  16. 0:28that. Hi, my name is Callum, also known
  17. 0:30as Waterlitz, and welcome to today's
  18. 0:31video on giving your agent true memory.
  19. 0:34Today I'm using Hermes agent as an
  20. 0:35example, but the memory tools I talk
  21. 0:37about apply to any different agentic AI.
  22. 0:39And more importantly, they work between
  23. 0:41any AI agents. You can set this system
  24. 0:43up once and then connect it to any agent
  25. 0:45with the same universal memory. If
  26. 0:47you're new to Hermes and want to get it
  27. 0:48set up safely, I have a full agentic AI
  28. 0:50playlist here. By the end of this video,
  29. 0:52you'll understand the different tiers of
  30. 0:54the memory stack and how agentic memory
  31. 0:55compares to a knowledge layer like
  32. 0:57Obsidian LLM Wiki. Why you might want a
  33. 0:59dedicated memory provider instead of
  34. 1:01using built-in memory, the different
  35. 1:02memory options and why I picked the two
  36. 1:04we're setting up today, how to install
  37. 1:05an unofficial lightweight memory layer,
  38. 1:07and how to install an official
  39. 1:09heavyweight memory engine. I'm
  40. 1:11personally very excited because memory
  41. 1:12is something I've been looking at for a
  42. 1:13long time, and I can't wait to help you
  43. 1:15get it set up today. Now, let's take a
  44. 1:17look at agentic memory.
  45. 1:22Before we get anything installed, I just
  46. 1:24want to quickly clear up some common
  47. 1:25confusions on agentic memory. It's not a
  48. 1:28layer, it's a stack. When people say
  49. 1:30second brain or agentic memory, they're
  50. 1:32actually referring to three things:
  51. 1:34world knowledge, built-in memory, and
  52. 1:36external memory. The knowledge layer
  53. 1:38like Obsidian vaults or an LLM Wiki is
  54. 1:40part of the memory stack, but it's not
  55. 1:41true operational memory the way we mean
  56. 1:43it when we say agent memory. It's world
  57. 1:45knowledge, the ground truth across every
  58. 1:47project and every agent. You can even
  59. 1:49limit the access your agent has to this
  60. 1:51knowledge layer, like I talked more
  61. 1:52about in connecting agentic AI to
  62. 1:54Obsidian here. Then we have operational
  63. 1:56memory, which splits into two tiers,
  64. 1:58built-in memory and dedicated memory
  65. 2:00providers. Operational memory focuses
  66. 2:02more on what the agent should remember
  67. 2:03about working with me specifically. A
  68. 2:05simple test is to consider if it's
  69. 2:07knowledge you want preserved and
  70. 2:08organized across everything you do,
  71. 2:10that's Obsidian. If it's a fact about
  72. 2:12how you like to work or what you're
  73. 2:13working on right now, that's agentic
  74. 2:15memory. So, putting it all together, we
  75. 2:17have built-in memory, which is great for
  76. 2:18simple single session chats, a memory
  77. 2:20provider when you're running complex
  78. 2:22ongoing projects or you just want
  79. 2:23noticeably better performance and agent
  80. 2:25recall, and your knowledge layer, like
  81. 2:27Obsidian, for building world knowledge
  82. 2:29that spans every project and every agent
  83. 2:31you run. I talked more about getting
  84. 2:32Obsidian connected to agentic AI in this
  85. 2:34video here. So, today I want to focus
  86. 2:36more specifically on the first two
  87. 2:37layers, on the built-in memory and the
  88. 2:39external memory provider. My next video
  89. 2:42is going to go deeper into setting up an
  90. 2:43Obsidian LLM wiki that complements the
  91. 2:45agent memory we're setting up today.
  92. 2:47Let's start by comparing the built-in
  93. 2:48memory to true memory providers. If you
  94. 2:50find this video helpful, please like,
  95. 2:52hype, and subscribe as I appreciate your
  96. 2:53support a lot. If you're looking for
  97. 2:55more ways to support me, please consider
  98. 2:56joining my YouTube or my Patreon
  99. 2:57membership, where I give my members
  100. 2:59access to tips, insights, and my
  101. 3:01thoughts on the future of agentic AI and
  102. 3:03knowledge management.
  103. 3:08So, what does built-in memory actually
  104. 3:09look like for Hermes? It has three
  105. 3:11files, memory.md, which is your
  106. 3:13operational memory, user.md, which is
  107. 3:15who you are, and soul.md, which is
  108. 3:18technically not memory but the agent's
  109. 3:19own personality. On top of that, the
  110. 3:21built-in memory includes session search,
  111. 3:23stored in a local database, so you can
  112. 3:25literally ask Hermes to search across
  113. 3:27your past sessions. As a quick tip, the
  114. 3:29best way to get going with an agent like
  115. 3:31Hermes is to tell Hermes about yourself
  116. 3:33right off the bat. If it understands a
  117. 3:35little bit more about who you are and
  118. 3:37how you like to work, the whole system
  119. 3:38gets so much smarter almost immediately.
  120. 3:41One question I get a lot about running
  121. 3:42the built-in memory is if you're using
  122. 3:44Hermes terminal execution back end,
  123. 3:46you're running the code in something
  124. 3:48like Docker, can Hermes still access the
  125. 3:50memory? And the answer is yes, it can.
  126. 3:52Running in Docker doesn't impact Hermes'
  127. 3:54ability to access its own memory and
  128. 3:56user files. That's done using a memory
  129. 3:58tool, which is the same thing that the
  130. 4:00memory providers use, rather than
  131. 4:01running through code that ends up being
  132. 4:03sandboxed in Docker. I talked more about
  133. 4:05that in my how to get your agent
  134. 4:07isolated and operating more safely on
  135. 4:09your computer. The key is that the
  136. 4:10built-in user and memory files get
  137. 4:12injected into every new session. It
  138. 4:14becomes part of the context of every
  139. 4:16chat you have with Hermes. Now, this is
  140. 4:18great because it means that your agent
  141. 4:19never forgets who you are, but it also
  142. 4:21means that it's using up some of the
  143. 4:22context and tokens for every single
  144. 4:24chat. You want to keep this as lean as
  145. 4:26possible, so you're not wasting tokens.
  146. 4:28And that's where having an actual memory
  147. 4:29provider comes in. A memory provider is
  148. 4:31an external layer that stores and
  149. 4:32retrieves facts at runtime or when the
  150. 4:34agent actually needs it, rather than
  151. 4:36inserting it into every conversation.
  152. 4:38This helps keep your context lean and
  153. 4:40your facts on point for when you
  154. 4:41actually need them. Let's take a deeper
  155. 4:43look into memory providers.
  156. 4:48A lot of people feel overwhelmed when
  157. 4:50trying to pick a memory provider, and
  158. 4:52that's fair. There's a lot of options
  159. 4:53and things to consider. It depends on
  160. 4:55your goals, how many agents or people
  161. 4:57are accessing that shared memory, what
  162. 4:59hardware you have available to run it,
  163. 5:00and whether you want to use free or paid
  164. 5:02tools. The good news is that most, if
  165. 5:04not all of them, allow you to migrate
  166. 5:05between each other, so you're not locked
  167. 5:07into one that you pick up front. You can
  168. 5:09test and experiment, and if you don't
  169. 5:11like one tool, you can always migrate
  170. 5:13your memories to another one. But rather
  171. 5:14than comparing them all in this video,
  172. 5:16which would take a while, I put together
  173. 5:17my analysis free Patreon post that I'll
  174. 5:20link in the top comment of the video, so
  175. 5:22you can take a look at the different
  176. 5:23options and see what works best for your
  177. 5:24scenario. But today, rather than going
  178. 5:26off of benchmarks or feature lists, I
  179. 5:28thought I would take a look at what the
  180. 5:29community has been rating the most
  181. 5:30highly and what many people have
  182. 5:32recommended to me. And they are Nemesis
  183. 5:34and Hindsight. Nemesis is a
  184. 5:35zero-dependency lightweight system and
  185. 5:37doesn't require an LLM to work. It runs
  186. 5:40fully locally, incredibly fast with
  187. 5:42built-in embeddings. Nemesis is great if
  188. 5:43you want a memory provider that doesn't
  189. 5:45slow your system down. Hindsight has
  190. 5:47more features and it ranks at or near
  191. 5:48the top of most memory benchmarks, but
  192. 5:50it either needs a cloud connection or
  193. 5:52your own server plus an LLM to actually
  194. 5:54power it. The simplest way to think of
  195. 5:56it is Nomasyne is a memory layer,
  196. 5:58Hindsight is a memory engine. So, let's
  197. 6:00start with the lightweight option of
  198. 6:01Nomasyne so you can see a memory
  199. 6:02provider in action before we move on to
  200. 6:05a little more power with Hindsight.
  201. 6:11So, we can think of Nomasyne as a
  202. 6:12universal memory layer for any AI agent
  203. 6:15running a SQLite backend with
  204. 6:17sub-millisecond zero dependencies. So,
  205. 6:19the purpose of showing you this one
  206. 6:21first is that it is very fast and it
  207. 6:24doesn't require you to install a lot of
  208. 6:26other things. So, a little bit more
  209. 6:27about Nomasyne is that it's biologically
  210. 6:30inspired by the BEAM architecture, which
  211. 6:32brings in different memory tiers like
  212. 6:33working, episodic, semantic, and a
  213. 6:35scratchpad where each one is optimized
  214. 6:38for different access patterns.
  215. 6:39Everything stays local, which is good
  216. 6:41for privacy. It has sub-100 millisecond
  217. 6:43queries and it has native Hermes
  218. 6:45integration.
  219. 6:49Let's take a look at the architecture
  220. 6:50for a moment. So, the key is that we're
  221. 6:52building auto context injection, which
  222. 6:54basically means that when you have a
  223. 6:56query, when you ask Hermes something,
  224. 6:59Nomasyne will receive that query, use a
  225. 7:02recall tool to go take a look at its
  226. 7:04existing context, and inject that
  227. 7:06context into the agent's prompts before
  228. 7:09it responds to you. As you continue
  229. 7:11having conversations, it's going to
  230. 7:12remember more, which is going to get fed
  231. 7:14back into Nomasyne to help it build this
  232. 7:16context assembly. So, this is just one
  233. 7:18great way to continue building out a
  234. 7:20self-evolving memory system for your
  235. 7:21agent. And like we can see, it can be
  236. 7:23used for Claude code, for Codex, it can
  237. 7:25be used for Cursor, it can be used for
  238. 7:27any agent that you want, and we can have
  239. 7:29all of them connect to the same memory
  240. 7:31bank if we want to. All right, so let's
  241. 7:33get Nomasyne set up.
  242. 7:37So, if we go to their docs here, we can
  243. 7:38see that they have Hermes agent
  244. 7:39integration and it gives a list of quick
  245. 7:41commands here, so it's pretty easy to
  246. 7:43install. And one thing I want to mention
  247. 7:45just very quickly is that if you're
  248. 7:46running Hermes and you have your
  249. 7:47terminal execution back end running in
  250. 7:49Docker, we can't just give Hermes the
  251. 7:51task of setting up Mnemosyne for us
  252. 7:53because Hermes would try and run its
  253. 7:55terminal, and it would run inside of a
  254. 7:57Docker container here, and that's not
  255. 7:59what we need because we need our memory
  256. 8:01system to survive outside of the
  257. 8:03terminal back end Docker container. If
  258. 8:05you're not running inside a Docker, you
  259. 8:06can just ask Hermes to do it, but
  260. 8:07basically this just means that we need
  261. 8:09to run it from our host device inside of
  262. 8:11our computer. So, I'm going to install
  263. 8:12it directly from terminal myself.
  264. 8:14Mnemosyne is not a built-in provider, so
  265. 8:17we need to install Mnemosyne a little
  266. 8:19bit more manually. But the key here is
  267. 8:21that we're not installing it inside of
  268. 8:23Hermes managed virtual environment
  269. 8:26because every time you run Hermes
  270. 8:27update, it rebuilds that virtual
  271. 8:29environment, and then it wipes any extra
  272. 8:31packages. So, we have two options that
  273. 8:33are listed here. We can use pipx, which
  274. 8:35creates a separate virtual environment,
  275. 8:37or we can manually create a virtual
  276. 8:40environment using the fallback option.
  277. 8:41However, neither of these options worked
  278. 8:43perfectly for me. So, if you install
  279. 8:45Mnemosyne and you run Hermes memory
  280. 8:48status and you get a plugin missing
  281. 8:50error, or when you first install it, the
  282. 8:52Python environments don't match error,
  283. 8:54you might need to add another step in
  284. 8:56between.
  285. 8:59So, the first step is going to be to
  286. 9:01install Mnemosyne Hermes into its own
  287. 9:03pipx environment, then set the config.
  288. 9:06So, these two steps are the same, but
  289. 9:07then before we restart the gateway, we
  290. 9:09need to link the new environment that we
  291. 9:11just created into the Hermes plugins
  292. 9:14folder, and then we need to run what's
  293. 9:15called the Hermes memory setup. That's
  294. 9:18just a very easy way that you can run it
  295. 9:19inside of terminal, and it will go
  296. 9:21through and configure Mnemosyne for you
  297. 9:23so that it gets recognized as a memory
  298. 9:24provider instead of just as a plugin.
  299. 9:26So, let's get started. Okay, so I'm
  300. 9:27going to go to the documentation. I'm
  301. 9:29going to copy the install. I'm going to
  302. 9:31go to my terminal and paste this in.
  303. 9:33Great, there we go. That took less than
  304. 9:35a minute. So, we now have Mnemosyne
  305. 9:36Hermes set up. Now, if we go over to
  306. 9:38memory and context, we can see that we
  307. 9:40currently have the default memory system
  308. 9:42enabled, and this is the built-in memory
  309. 9:44provider. But, we can see here, if we
  310. 9:45click the drop-down, we don't have
  311. 9:47access to Mnemosyne inside of here.
  312. 9:49That's because it's a plugin, it's not a
  313. 9:52built-in memory provider. So, we need to
  314. 9:54enable that for our default profile
  315. 9:56outside of the settings of the desktop
  316. 9:58app. So, let's do that for a moment. So,
  317. 9:59I'm going to go back to the docs here,
  318. 10:02copy this string, paste it in. Great,
  319. 10:04there we go. That just took 2 seconds.
  320. 10:06And basically, what just happened there
  321. 10:07is it modified the config.yaml file to
  322. 10:10use Mnemosyne as the memory provider.
  323. 10:13Now, we just installed Mnemosyne into
  324. 10:16One Bin, and we want to create a symlink
  325. 10:19or a link between where we just
  326. 10:21installed it using that separate
  327. 10:22environment and the actual Hermes plugin
  328. 10:25folder right here. So, that's what we
  329. 10:26can see right here. It just went through
  330. 10:28and verified that the provider of
  331. 10:30Mnemosyne is available and that it was
  332. 10:32able to connect Mnemosyne to Hermes. So,
  333. 10:34our next step is to run Hermes memory
  334. 10:36setup, and this is where we can see
  335. 10:37here, we want to go to Mnemosyne local,
  336. 10:40click yes. So, now that we are running
  337. 10:42Hermes memory setup, we're going to go
  338. 10:44through step-by-step and select all of
  339. 10:46the different configurations we want for
  340. 10:48Mnemosyne as Hermes memory provider. I
  341. 10:50just used the default option for
  342. 10:52everything except for being able to
  343. 10:54access global rather than session
  344. 10:56memories. But, I do recommend reading
  345. 10:58what it's asking, and if you're not
  346. 10:59sure, asking Hermes to help you
  347. 11:01understand the setup in a way that's
  348. 11:02going to work best for you. The most
  349. 11:04important setting, what's the default
  350. 11:05scope for remembering? I don't want to
  351. 11:07limit it per session, I want to have it
  352. 11:10be across all sessions. That's one of
  353. 11:11the benefits of running a memory system,
  354. 11:13so it's really important to switch this
  355. 11:15one to global. And there we go. So, we
  356. 11:17can see that this just went through and
  357. 11:18it saved the memory provider of
  358. 11:20Mnemosyne. So, now what we do is, if we
  359. 11:22go over to config.yaml inside of our
  360. 11:24.hermes folder, there we go. We can see
  361. 11:26that we have a whole memory system here
  362. 11:28with all of these different skills and
  363. 11:30tools set up for the Mnemosyne provider.
  364. 11:32So now I'm going to restart the Hermes
  365. 11:34Gateway. Going to type in Hermes memory
  366. 11:36status and we can see that the memory
  367. 11:38tools enabled, the provider is Nemesis,
  368. 11:40we have all of the configurations that
  369. 11:41we just went through so we can change
  370. 11:43this later if we want. We have the
  371. 11:44plugin installed and the status is
  372. 11:46available which is key. And now we can
  373. 11:48see here if I type in Hermes Nemesis
  374. 11:50stats, everything is currently at zero
  375. 11:52because we haven't actually stored
  376. 11:53anything yet. So let's get that started.
  377. 11:58I'm going to open the Hermes desktop app
  378. 12:00and if you go over to memory and
  379. 12:01context, we can see that Nemesis is now
  380. 12:03set up as the memory provider. Great. So
  381. 12:05now we have it all installed and if we
  382. 12:07ever want to upgrade this, we can just
  383. 12:08type in pipx upgrade Nemesis-Hermes.
  384. 12:11One more step I want to quickly mention
  385. 12:13here is that the Nemesis documentation
  386. 12:15suggests disabling legacy memory. And
  387. 12:17basically what that means is just this
  388. 12:19persistent memory and user profile we
  389. 12:21have set. The main reason is to avoid
  390. 12:23duplication and token waste. So I would
  391. 12:25suggest disabling it to start and then
  392. 12:28testing it out and if you notice a
  393. 12:29problem, you can always re-enable it.
  394. 12:31But don't disable it using a command
  395. 12:33like this. Make sure you just update it
  396. 12:35inside of the config.yaml file where we
  397. 12:38can see memory enabled false, user
  398. 12:39profile enabled false, which is the same
  399. 12:42thing as toggling these buttons off
  400. 12:44right here.
  401. 12:47Okay, so let's quickly test it out. If I
  402. 12:49ask Hermes a question, it shouldn't know
  403. 12:51anything about me. So previously I had
  404. 12:53told Hermes who I was, but now why don't
  405. 12:55we try it again since we have Nemesis
  406. 12:57set up. So we can see here it's going
  407. 12:58and it's actually enabling the Nemesis
  408. 13:00recall skills and the canonical skill,
  409. 13:03but it doesn't have anything. So I'm
  410. 13:05going to give it a little bit about me
  411. 13:06so that it can remember it. Just as a
  412. 13:07quick tip, this is always the first
  413. 13:09thing I would do when you're setting up
  414. 13:10a new memory system is give Hermes or
  415. 13:12your agent some context on who you are
  416. 13:14and what you like. That way it starts to
  417. 13:16build a profile around you and the chat
  418. 13:19will start to feel more personalized and
  419. 13:21effective for what you're trying to
  420. 13:22accomplish. So my name is Callum, I go
  421. 13:24by Wander Roots, I run a YouTube
  422. 13:26channel, I'm a patent agent,
  423. 13:27intellectual property lawyer. I studied
  424. 13:29mechanical materials science
  425. 13:30engineering, and this is my preference
  426. 13:32for how I like to respond. So, we can
  427. 13:35see here, now Hermes is going through
  428. 13:37and is running the Nemesis tool. It's
  429. 13:39running remember. It has professional
  430. 13:41background, alias, and name. And all of
  431. 13:44these things are being stored into the
  432. 13:47database that Nemesis has set up. So,
  433. 13:50now for example, if I open a new chat
  434. 13:51and I say who am I, we should see Hermes
  435. 13:54recall from Nemesis. Here we go. It
  436. 13:56pulled it in. That's awesome. That means
  437. 13:57that it is working perfectly. And now if
  438. 13:59we go back to terminal and run Hermes
  439. 14:01Nemesis stats, we can see here there's a
  440. 14:02total of four pieces of working memory
  441. 14:05unconsolidated. So, this is where it's
  442. 14:07going to go through and over time it's
  443. 14:08going to shift working memory into
  444. 14:10episodic memory and then deeper into it.
  445. 14:12And this is bringing in what's called
  446. 14:14the beam architecture. So, it's going to
  447. 14:16observe, it's going to take in
  448. 14:17conversation and tool results, and it's
  449. 14:19going to remember all of these things
  450. 14:20into its working memory. Then over time
  451. 14:22it's going to consolidate that into its
  452. 14:24long-term experiences or its episodic
  453. 14:26memories. Every time we ask the agent
  454. 14:28something, every time there's an input,
  455. 14:30it's going to go through and pull from
  456. 14:31both of those and then recall it. So,
  457. 14:33there's more details on working memory
  458. 14:36but this is a pretty cool system
  459. 14:37especially considering that it's all
  460. 14:39happening locally inside that SQLite
  461. 14:41database. And what's cool too is it's
  462. 14:43also bringing in structured knowledge
  463. 14:45stored as subject predicate object
  464. 14:47triples which is the basis of a
  465. 14:48knowledge graph. And if you want to
  466. 14:50learn more about that, I do have a video
  467. 14:51on knowledge graphs and how we can use
  468. 14:53it for improving agent systems. But I
  469. 14:55highly recommend going through and just
  470. 14:56exploring the beam overview so you
  471. 14:58understand a little bit more how this is
  472. 15:00all operating. This is pretty cool.
  473. 15:05Remember, if we go here, we had to
  474. 15:06manually add Nemesis to this list. But
  475. 15:09there's all these built-in ones that
  476. 15:11perhaps are a little more powerful
  477. 15:12depending on what you're looking for.
  478. 15:14And I'm going to work on installing
  479. 15:15hindsight today. So, I personally like
  480. 15:18the fact that Nemesis can run locally,
  481. 15:21it runs without me needing to create
  482. 15:24anything extra. It doesn't have any
  483. 15:25dependencies and it's very fast. So, I'm
  484. 15:27going to leave this as my default right
  485. 15:28now, but I'm going to create a new
  486. 15:29profile and I'm going to call it
  487. 15:31hindsight and I'm going to use this to
  488. 15:33set up a separate configuration here.
  489. 15:35Okay, so I'm now on the hindsight
  490. 15:36profile and if we go over to memory and
  491. 15:38context again, we can see here that I
  492. 15:40have persistent memory and user profile
  493. 15:42still turned on and the built-in memory
  494. 15:44provider is set. So, I don't have this
  495. 15:47one set up to connect to Namaseni. You
  496. 15:50can see that the hindsight profile
  497. 15:51doesn't even see it here. So, what we
  498. 15:53can do is we can go over to memory
  499. 15:55provider in our settings. We can go down
  500. 15:57to hindsight and turn this on. But, just
  501. 15:59before I do that, let's take a look at
  502. 16:00the hindsight website itself.
  503. 16:06So, hindsight is a native memory
  504. 16:08provider in the Hermes agent, which
  505. 16:10means it's built-in. And hindsight is a
  506. 16:12little bit different because it operates
  507. 16:14by having the agent actually connect to
  508. 16:16a separate server. So, it's not
  509. 16:18something that's running within Hermes
  510. 16:19itself, but what's cool is it has a
  511. 16:21whole bunch of different ways of
  512. 16:23operating. It builds its own mental
  513. 16:25models. It pulls in observations. It has
  514. 16:28standard memories. There's a lot you can
  515. 16:29do here and it can even reflect and then
  516. 16:32it's going to go through and extract
  517. 16:34more aware responses using a deeper
  518. 16:37sense of reasoning. So, when you run
  519. 16:39reflect, it runs an agentic loop that
  520. 16:41automatically searches the memory and
  521. 16:43then applies that to shape its own
  522. 16:44reasoning style and then produces a
  523. 16:47final answer grounded in what it's
  524. 16:48found. So, rather than just returning
  525. 16:50raw facts, reflect is a more synthesized
  526. 16:53response, which is why it also requires
  527. 16:55an LLM. So, there's a little bit more
  528. 16:57complexity here. I highly recommend
  529. 16:59exploring the hindsight documents if you
  530. 17:01want to learn a little bit more about
  531. 17:02it, but the key here is that this
  532. 17:04requires actually running a separate
  533. 17:06server or connecting to a separate
  534. 17:07server. So, if we go back to our memory
  535. 17:09provider here and select hindsight, we
  536. 17:12can see that a mode appears.
  537. 17:16So, this mode we have the option to
  538. 17:18switch between cloud and local external.
  539. 17:20And to understand that a little bit
  540. 17:22more, we can go specifically to the
  541. 17:23Hindsight docs on setting up for Hermes
  542. 17:25desktop specifically. And if you want to
  543. 17:27do it for the command line, there's also
  544. 17:29a separate way to get it going using the
  545. 17:30Hermes memory setup like you saw we did
  546. 17:33for Namacini. So basically we just need
  547. 17:34to select Hindsight, set what mode we
  548. 17:36want. And here we have cloud which
  549. 17:39requires an API key. That's where we can
  550. 17:41connect to the Hindsight cloud. But this
  551. 17:43does have a fee to it. It has you go so
  552. 17:47you can start for free and then
  553. 17:48depending on how often you're using it,
  554. 17:50depending on how big your system gets,
  555. 17:52you are going to have to start paying
  556. 17:53for it. You could just run the system
  557. 17:55yourself for free on your own computer
  558. 17:57if you have the hardware to do so. But
  559. 17:59if you don't, it might be worth taking a
  560. 18:00look at Hindsight cloud to see how much
  561. 18:02it costs. I personally haven't tried
  562. 18:03this yet. And today we're going to set
  563. 18:05up a local system. So take a look at
  564. 18:07this comparison here on comparing local
  565. 18:10to cloud LLMs for running your agent
  566. 18:11memory. So this is the LLM that's going
  567. 18:14to power the actual system as opposed to
  568. 18:16connecting to Hindsight cloud which also
  569. 18:18deals with storage. You can just connect
  570. 18:20to an open AI Anthropic or Gemini API,
  571. 18:22but then you're going to pay per token
  572. 18:24for every time you retain it. So you can
  573. 18:26run free locally on your own computer
  574. 18:28using something like Ollama which I have
  575. 18:30a video on running Ollama for agentic AI
  576. 18:32if you're interested in seeing how that
  577. 18:33works. One thing to keep in mind is that
  578. 18:35not all local models support tool
  579. 18:37calling. So you have to make sure that
  580. 18:38the one you're using does. I'm going to
  581. 18:40use GPT-OSS 20 billion parameters which
  582. 18:43is 13 gigs. So make sure you have that
  583. 18:45model installed already. And also make
  584. 18:47sure that you have the Ollama app
  585. 18:49running in the background or serve it
  586. 18:51yourself from terminal.
  587. 18:55So our next step now that we have the
  588. 18:57local model downloaded is that we need
  589. 18:58to connect Hindsight to that local model
  590. 19:02and get the server running in the
  591. 19:04background. So I'm going to do this
  592. 19:05inside of a docker container because
  593. 19:07this just improves the safety. Right now
  594. 19:09I already have my Hermes terminal back
  595. 19:10end is connecting to a docker container,
  596. 19:12so we're going to add another Docker
  597. 19:14container here that runs our memory
  598. 19:16system in hindsight directly. We can
  599. 19:18paste in this string, which I'll include
  600. 19:19in the free article linked in the top
  601. 19:21comment, but basically we want to run
  602. 19:23the Docker container with the name of
  603. 19:25hindsight, and we want it to even if the
  604. 19:26computer shuts down and starts up again,
  605. 19:28we want it to restart on its own. We're
  606. 19:30giving it two different ports, one for
  607. 19:32running the actual server, and then one
  608. 19:33for us to be able to look at the
  609. 19:34dashboard. And then these are all just
  610. 19:37environment variables that we are
  611. 19:39configuring so that it can connect to
  612. 19:41specifically the model that I have just
  613. 19:43selected here, which is the GPT-OSS 20
  614. 19:45billion. Okay, there we go, and it was
  615. 19:47able to download hindsight for us. And
  616. 19:49now if we go over to Docker Desktop,
  617. 19:51which by the way this has to be running,
  618. 19:52I have a whole video that goes deeper
  619. 19:54into how to get your Hermes agent and
  620. 19:56now something like hindsight
  621. 19:57containerized inside of Docker, but we
  622. 19:58can go here and we can see that we have
  623. 20:00our server is now running. So if we go
  624. 20:02over to dashboard here, we can select
  625. 20:04memory bank, we can create a new one, we
  626. 20:06can get everything going. But this just
  627. 20:08shows us that we have the server running
  628. 20:10and it's available for us to connect to.
  629. 20:12Okay, so our next step is to connect
  630. 20:14this running server to Hermes. Let's do
  631. 20:16that for a moment.
  632. 20:20So remember I'm in my hindsight profile,
  633. 20:22I'm going over to memory and context. We
  634. 20:24have hindsight settings here since we've
  635. 20:26selected the drop-down, and I'm going to
  636. 20:27change this to local external. Because
  637. 20:29of that we can leave the API key blank.
  638. 20:31I'm going to instead of having this
  639. 20:32connect to hindsight cloud, I'm going to
  640. 20:34put in the local host server that we
  641. 20:36just set up. I'm going to give it the
  642. 20:37name of Hermes, and I'm going to keep
  643. 20:39this recall budget as mid. So this isn't
  644. 20:42doing anything to control the AI model,
  645. 20:44this is just telling Hermes where to
  646. 20:45find the hindsight program that we just
  647. 20:48launched. So now just to test this, I'm
  648. 20:49going to go back and I'm going to
  649. 20:51actually turn off persistent memory.
  650. 20:52This is something that you can run in
  651. 20:54the background with hindsight. I'm just
  652. 20:56going to turn it off for a second so we
  653. 20:57can test hindsight. And this is
  654. 20:58different than Namaste, which recommends
  655. 21:01disabling this. So I'm going to say who
  656. 21:02am I? There we go. So we can see
  657. 21:04hindsight recall just triggered, but it
  658. 21:06doesn't know who I am. So, I'm giving it
  659. 21:08the same prompt as before to see if it
  660. 21:10can remember this. We can see it's
  661. 21:12planning durable memory. It's trying to
  662. 21:14update it, but again, in memories I
  663. 21:16turned that off. So, now it's running
  664. 21:18hindsight retain, and this can take a
  665. 21:20little bit because I'm running a local
  666. 21:21model. But, if we go take a look at
  667. 21:23Docker for a second, we can go to my
  668. 21:24hindsight container. We can see here
  669. 21:26that it's running triggering the memory
  670. 21:28engine. So, if the model that I selected
  671. 21:30has gone to sleep, it can take a little
  672. 21:31bit. But, we can see here it's working
  673. 21:33to store in chunks. Okay, there we go.
  674. 21:35It was able to use hindsight retain to
  675. 21:38store the user profile and preferences.
  676. 21:40Cool. So, now it's saying too, the most
  677. 21:41useful optional context would be what
  678. 21:43are the answers to all of these things.
  679. 21:45But, I'm just going to start a new
  680. 21:46session and say, "Who am I?" and see if
  681. 21:48it's able to pull that information from
  682. 21:50hindsight. There we go. So, rather than
  683. 21:51retain, it's now running hindsight
  684. 21:53recall, and it pulls up all the details
  685. 21:55that we had before, which is great. That
  686. 21:57means that this memory provider is
  687. 21:59working.
  688. 22:02And now, if we go back to localhost 99
  689. 22:05dashboard, we can select the Hermes
  690. 22:07memory bank that now exists here. We can
  691. 22:09click on it, and what's super cool is we
  692. 22:10have the ability to see all of the
  693. 22:12information here that's already been
  694. 22:14created as part of the hindsight system.
  695. 22:17So, to me this is pretty wild that not
  696. 22:19only are we able to get the agent to
  697. 22:20remember things, but it's able to build
  698. 22:22this graph based on our memory, and
  699. 22:24there's so much detail we can get into
  700. 22:26here. This is called constellation view.
  701. 22:28We can switch it to table. We can take a
  702. 22:30look at timeline, which would happen as
  703. 22:32we get more details. There's
  704. 22:33experiences, observations, mental
  705. 22:35models. There's so many things that we
  706. 22:38can structure here that I'm personally
  707. 22:39super excited to dive into.
  708. 22:44So, for example, I'm just going to
  709. 22:45answer the question that Hermes had
  710. 22:47given me, giving it a little more
  711. 22:48context on what I hope to do with my
  712. 22:51YouTube channel, who I like to help, why
  713. 22:53I'm doing this in the first place. So,
  714. 22:55I'm going to have hindsight now retain
  715. 22:57this information. Okay, so it was able
  716. 22:59to run there, and if I click refresh on
  717. 23:01this graph, we can see that there's now
  718. 23:03a few more pieces of information. I can
  719. 23:05tell Hermes to please reflect on this
  720. 23:06conversation.
  721. 23:09So, here's the result of the reflection
  722. 23:11that also pulled in some recall. There
  723. 23:13was a bit of an issue, but it seems to
  724. 23:14have done a pretty good job pulling in a
  725. 23:17deeper understanding of what it is I'm
  726. 23:19trying to do just based on that brief
  727. 23:21discussion I've had. I'm pretty
  728. 23:22impressed that it's pulling this in so
  729. 23:24quickly. The capable but overloaded
  730. 23:26builder people that I'm chatting with
  731. 23:28and I'm teaching here are curious and
  732. 23:30are wanting to understand rather than
  733. 23:32just getting answers, but I hope you can
  734. 23:33see that not only are these memories
  735. 23:35being stored here, they're being used in
  736. 23:37a way that improves the reflection that
  737. 23:39the agent has on the problems I'm trying
  738. 23:41to solve and what I'm interested in
  739. 23:43doing. And that was just after a basic
  740. 23:45amount of information. So, that's pretty
  741. 23:47cool. And again, this reflection is
  742. 23:48something that from what I understand at
  743. 23:50least at this moment, Mnemosyne doesn't
  744. 23:52have yet. So, that brings me to the
  745. 23:54question of which one should you choose?
  746. 23:56What should we actually care about here
  747. 23:59and where should we take this next?
  748. 24:05The author of Mnemosyne says these are
  749. 24:07not direct competitors. Hindsight is a
  750. 24:09memory engine with sophisticated natural
  751. 24:11language processing and multi-signal
  752. 24:13retrieval. Mnemosyne is a memory layer
  753. 24:15optimized for simplicity, speed, and
  754. 24:17single machine deployments. So, this is
  755. 24:19really the summary of it. Hindsight is
  756. 24:21an engine, whereas Mnemosyne is a layer.
  757. 24:24So, it's up to you to decide what works
  758. 24:25best for you. I am going to put more
  759. 24:27information here and all of the
  760. 24:28resources I talk about today into a free
  761. 24:30Patreon article that will go more
  762. 24:32in-depth on how to make this decision
  763. 24:34yourself, think the key here is to test
  764. 24:36one provider for a week and see if it
  765. 24:38improves your workflow. If it doesn't,
  766. 24:40try another one. You can always export
  767. 24:42from one to the other, there's no
  768. 24:43lock-in. And again, one thing that I
  769. 24:45personally am excited to explore a lot
  770. 24:47more is the dashboard here. I just think
  771. 24:49this is so cool that I'm able to go and
  772. 24:51see all of these elements coming
  773. 24:53together, this full dashboard of the
  774. 24:55memories that are being created from my
  775. 24:57agent as I just have conversations with
  776. 24:59it. And to me, this is also the key
  777. 25:02benefit of running something like
  778. 25:03hindsight over Nemesimy because while
  779. 25:06Nemesimy does have a dashboard, it's a
  780. 25:09community dashboard. It's not built by
  781. 25:10the creator of Nemesimy. So, Nemesimy is
  782. 25:13a smaller tool built by a smaller team
  783. 25:15and hindsight is a full system that's
  784. 25:18got a lot more going for it. And that
  785. 25:20brings me to the next step, which I'm
  786. 25:21really excited about. Not just building
  787. 25:23memories based on my chats with the
  788. 25:25agent, but also connecting this to my
  789. 25:26world knowledge, my library, or my
  790. 25:29second brain inside of my Obsidian LLM
  791. 25:31Wiki.
  792. 25:34So, now your agent has true agentic
  793. 25:36memory. Nemesimy gives you a fast
  794. 25:38lightweight memory layer, and if you're
  795. 25:39looking for a little more power, you can
  796. 25:41use the memory engine of hindsight. If
  797. 25:43you have any questions about what we
  798. 25:44talked about today or tips for how
  799. 25:45you're using your own workflow, please
  800. 25:47let me know in the comments. I'm happy
  801. 25:48to answer questions and I love learning
  802. 25:50how you're using these systems as well.
  803. 25:52But, there's still one piece missing,
  804. 25:53world knowledge. Your operational
  805. 25:55memory, the agentic memory we set up,
  806. 25:57now knows how you work. Next, I'll
  807. 25:58connect it to the Obsidian LLM Wiki so
  808. 26:00your agent isn't just remembering facts
  809. 26:02about you. It's also reasoning against
  810. 26:04everything you already know. If you
  811. 26:06found this video helpful, please like
  812. 26:07and subscribe. I really appreciate it.
  813. 26:09And don't forget to check out the free
  814. 26:10Patreon post that I've linked below with
  815. 26:12more information on the setup, the code,
  816. 26:14and the memory provider options. Thanks
  817. 26:16again for watching and I will see you in
  818. 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.