Aditya Agarwal warns AI agent memory suffers from structural limits

Reactions from the X community

7:57 AM · Jul 21, 2026

The root cause is that we are using the chat transcript as continuation state, and it's really poor at that. It's ambiguous, and surface-bound. Better compaction and more complex memory systems aren't the answer, yet many are chasing that and just throwing more tokens at symptoms. As you suggest, there is a missing primitive, one of governed semantic state. This is something I'm working on.

KG Krishna Gade@krishnagade

8:37 AM · Jul 21, 2026

Agent memory is still one of the biggest unsolved problems. Today, we are able trace tool calls and model outputs, but not what the agent knew, forgot, compressed, or dropped along the way. Skills help with procedures, but they don’t really solve state. Need a real context layer with provenance, retention rules, and compaction history.

PS PJ Standley@ArkStabler

5:24 PM · Jul 21, 2026

I made a shared brain for my agents with @databricks They check in end of day and beginning of the day to report on any movement on any of our projects. It’s still search plus context window management but it doesn’t feel like it now. it’s been a fun and fruitful Experiment.

MS Myles Spencer@mespercer

6:16 AM · Jul 21, 2026

@adityaag I’ve found having the main thread you’re interacting with plan, delegate to other agents, and verify (i.e not doing any actual work itself) helps this issue significantly. It doesn’t completely solve the issue, but makes it much more bearable

View all

DM Dave Morin 🦞@davemorin

5:08 PM · Jul 21, 2026

1214TECH

@adityaag Agent memory is one of the biggest problems in the game. I fear we may need to be training some subset of weights daily in order to make real progress on this. Otherwise it is all just some combination of search plus context window management.

MS Marco Sanvido@msanvido

12:11 PM · Jul 21, 2026

@adityaag promising: https://cloud.google.com/blog/products/data-analytics/how-the-open-knowledge-format-can-improve-data-sharing

AA Aditya Agarwal@adityaag

5:13 AM · Jul 21, 2026

1745TECH

Memory "loss" and compaction still feels like a huge issue with all harnesses. A tendency to forget and get confused is pretty frustrating for the end user. The "interpretability" when these errors occur is also pretty poor. Skills as a way to store this information feels like one of the root issues. Would be cool to have some kind of format/language to make this better.

Aditya Agarwal warns AI agent memory suffers from structural limits · Digg

Aditya Agarwal warns AI agent memory suffers from structural limits

Combined views

Aditya Agarwal warns AI agent memory suffers from structural limits

Reactions from ranked influencers

Combined views

Sentiment

Positive Read

Digg Deeper

Reactions from the X community

Sentiment

Positive Read

Digg Deeper