AI Agents can remember. But how does their memory actually work? 🧠 AI Agent Memory, explained in just 12 minutes! 🔥 • Working Memory & Context • Procedural, Semantic & Episodic Memory • RAG, Vector DBs & Top-K Search • Storing Past Conversations & Events • Memory Consolidation &... Summarization • Efficient Token Usage As AI agents take on bigger, longer tasks, memory will become essential to keep context, learn from experience, and build on past work.show more

Hello Vishal, I’d appreciate it if you tag my handle and the original link of the video.

hey Shen, already tagged you in the video....plus also DMed you before posting....but thanks for adding the youtube link here 👀🔥

Still it’s inappropriate to re-upload another creator’s videos imo. I think the tagging feature only allows me to see it.

nah tagging is visible to all checked 👍

@ShenSeanChen rag vector stuff still breaks half the time

@ShenSeanChen yup 😭

@ShenSeanChen Memory is the hard part. My agents remember my bugs better than I do 😅

@ShenSeanChen it's kinda the opposite for me nowadays my agents are acting way too dumb lol 😭

@ShenSeanChen Genius one day, goldfish the next. Agent life 😭

@ShenSeanChen the memory part is the whole game for agents. a model that forgets everything each session is just a very smart goldfish

@ShenSeanChen exactly 💯

@ShenSeanChen Memory is what turns a chatbot into an assistant. Without it you're re-explaining yourself every session.

@ShenSeanChen yup 💯

@ShenSeanChen 记忆机制讲得越来越细，用户这边还差个很实际的入口：它记错了，我去哪儿改？能搜到、能编辑、能删，长期记忆才让人放心。

@ShenSeanChen yes this point hasn't been covered here, thanks for pointing it out...need to check 👀👍💯

@ShenSeanChen the piece most memory setups skip is forgetting. if nothing ever expires, old wrong facts keep getting pulled in. a simple "last confirmed" date on each memory and a re-check when it's stale fixes a lot

@ShenSeanChen old wrong facts that are useless but still stored are the biggest hurdle...they fuck efficiency like no other 😭

@ShenSeanChen an agent that remembers last week starts to feel like the same agent every morning

@ShenSeanChen 🤨

@ShenSeanChen Episodic memory is the part people skip most. Agents keep facts fine but forget what they already tried. Do you get into how to decide what gets written to memory and what gets thrown away?

@ShenSeanChen kinda facing this my agents fail at complex tasks, so as i learn memory i'm trying to improve/optimise them better...will share more soon once i start seeing results 🫡

@ShenSeanChen For many SI agents, long-term memory lives in an external store. Retrieval brings selected memories into context, so the real test is finding the right detail when it matters

@ShenSeanChen yup 💯

@ShenSeanChen A useful memory test: store a project constraint, change it later, then ask the agent to plan against it. Does it use the correction, or retrieve both versions as equally valid? How are you handling superseded facts in your experiments?

@ShenSeanChen im experimenting this with something will share learnings/results soon 👀✨

@ShenSeanChen Would be interested to see the results. Another case worth testing: “use X for this task only” versus “replace X with Y from now on.” Can the agent keep a temporary exception from becoming a permanent project rule?

@ShenSeanChen Great breakdown 🔥 memory is the missing piece for agents doing real long-horizon work. Consolidation + summarization is where most implementations fall apart

@ShenSeanChen memory + evals, 2 very very imp things imo 🙌

@ShenSeanChen @vishalsingh2972 100% 🔥 bad memory write + no evals = agent confidently repeating the same mistake forever

@ShenSeanChen A memory system also needs a correction path. If a user changes a preference, can the agent supersede the old fact, avoid retrieving both as current, and explain where the new value came from? Remembering more is useful only if stale memories can lose authority.

@ShenSeanChen yes so much to needs to be done here...👀✨