All writing
·3 min·
AgentsMemoryProduct

Most AI agents remember too much

I gave one a five-tier memory architecture. It got confused. Here's what I learned.

I built an agent with a five-tier memory architecture. Episodic for what happened. Semantic for what it means. Procedural for how to act. Strategic for why it matters. Meta for how to learn.

On paper, it was impressive. In practice, it was confused.

It treated a three-week-old failed approach as valid input. It held onto context that had been explicitly overridden. It was loyal to the past in a way that made it unreliable in the present.

And I realized this is also a human problem. The best operators I've worked with don't have the best memories. They have the best filters. They know which signals to carry forward and which ones to drop.

Same principle applies to AI systems: forgetting on a schedule is a feature, not a bug. Recency bias, when intentional, is actually smart. The question 'what should this agent remember?' is a product decision, not just a technical one.

If you're building AI workflows for your team, ask: what outdated context is your system still acting on? It might not be a capability problem. It might be a memory hygiene problem.

Also published on, or originally from:

LinkedIn

LET'S TALK

Hiring? Or stuck mid-build?

Open to senior in-house roles with conversational AI / voice agent / AI-native DXP teams. Available for select Toptal engagements. I reply within 24 hours with something concrete — not a calendar link to a discovery deck.

→ Hiring conversations: tell me about the role and the stack

→ Project work: tell me the bottleneck, not the requirements doc

→ Either way, I'll come back with a take, not a pitch

LinkedIn DM

The fastest way to reach me. I'm already there every day — posts, comments, replies. Send a message and I'll come back within 24 hours.

/in/apsolut

GitHub

Pinned tools, side projects, the README that explains the vibe. CIA lands here when it's open-sourced.

/apsolut

Not ready to reach out?

Read what I've been writing about agents, model routing, and prompting. The fastest way to figure out if our wavelengths match.

Prefer email? Drop a line on LinkedIn first — I'll share the right address depending on whether the conversation is hiring, freelance, or something else.