Ex-Google engineer explains how to build memory for self-improving AI agents in 12 minutes. procedural (how to act / skills) + semantic (durable facts / profile) + episodic (dated events / chat history). Memory is an important part of an agentic loop. Memory + loops + harness + evals... = self-improving agent system. This is better than $500 paid courses on the internet - in under 12 minutes. Watch it today, then save the detailed guide on building agent memory from scratchshow more

creating a memory for self-improving AI agents is the best solution he has ever created

Yeah, that’s definitely a really important part of agentic loops.

free Anthropic courses and this == cheats

yeah, soo true !

Real talk

Yeah, real alpha from a senior engineer, real use cases, not just buzzwords.

this is a brilliant vid man

yeah bro, super useful for me too

The memory point is underrated. The useful version is not just “give Claude more notes,” it is deciding what becomes durable memory, what gets retrieved, and what should expire before it pollutes the next run.

@ShenSeanChen

Already bookmarked it, I absolutely love this kind of useful content, maybe I'll learn something myself

procedural + semantic + episodic is such a clean mental model for this

yeah bro, you nailed the right setup !

@0xCodez interesting breakdown! never thought about how memory types play a role in AI loops. makes me wonder how close we are to AI that can really adapt on its own.

the 3 memory types breakdown is genuinely solid agents without episodic memory are just goldfish with a system prompt

One challenge that doesn't get talked about enough is memory efficiency. As agents take on longer tasks, storing and retrieving the right information efficiently becomes increasingly important.

Splitting memory into procedural, semantic, and episodic is the part most agent tutorials skip, that structure is what actually makes recall useful instead of just a growing log file.

Encountered AtomicMemory and its philosophy resonates. Open source, self-hosted infrastructure for auditable memory. Rather than trusting closed systems, you maintain direct control and visibility.

actually, even though it's tedious, it's a very useful thing

episodic memory is underrated

The procedural/semantic/episodic split is the right taxonomy, and it's where most agent stacks are thinnest. Everyone bolts on a vector store for episodic recall and calls it memory. Procedural memory (learned skills, not just retrieved facts) is the part that actually compounds.

The biggest takeaway is that memory isn't just storing chat history. Different types of memory solve different problems.

episodic memory is the one people skip until context windows blow up. it's what scales session count.

The split matters because each type earns its keep differently. Episodic is the one most builds skip, and it is exactly what makes run 50 cheaper than run 1. One thing we learned shipping this: keep embeddings local. Recall gets so fast the model stops treating memory like an expensive last resort and starts calling it constantly, which is when agents actually start compounding.

brilliant share as always codez

Most teams get stuck on retrieval. The hard part is consolidation: what to retain, compress, or discard after each session. The summarizer loop is the actual product.

that architecture is exactly how we build smart agents

hit this exact wall with my finance app deprecated a skill in procedural memory but episodic wasn't triggering a rewrite so the old behavior kept firing for like 3 days before i caught it. did he cover invalidation logic, or just how to build the three layers?

built forge and i had to figure this out myself fr, memory is the hardest part

The taxonomy is the easy part. The hard part nobody demos in 12 minutes is consolidation: deciding what episodic noise gets promoted to durable semantic memory and what gets forgotten. Without that filter, memory just becomes a slower context window.

Nice framework. Memory + loops = powerful combo for smarter agents. Excited to see how this speeds up real-world progress.