Memory systems aren't personal ML
“Agent Memory” pitches are everywhere right now.
What’s missing in your agent’s brain is personal ML - training on YOU and YOUR project. Context engineering aims to solve that. Are memory systems the magic bullet?
The pitch is always some version of: your agent learns from experience, adapts to feedback, improves over time, no manual management. that’s personal ML. you don’t get personal ML from a sentence database, no matter how clever the pipeline.
A memory system is text. You give it text. Chunk it, map it, tag it, condense it. Hook it, update it. It still writes text.
An LLM is a neural network: weighted connections shaped by patterns across enormous amounts of text.
No single sentence lives inside it. what’s there is the residue of massive repetition.
No amount of engineering turns the 1st thing into the 2nd thing. A vector db of sentences, tagged, mapped, whatever, is still not a smaller version of an LLM.
Why does this matter for your Agent?
Example from a representative pitch: User says they’re an ML engineer. next message: “I just got promoted to CEO!” They say the mem system magically extracts, compares, judges how to update the old memory.
Sounds smart, but it might actually be an illustration of the problem. It’s making a permanent change based on one sentence.
Is that sentence true? A hypothetical? Is the person correct? In the correct sequence? What’s the context? A human would pause, ask, weigh, suspend judgment on incomplete signal.
A mem system can’t do that without a prompt for every edge case and massive context. It misses things, and it still encodes what will be, in the end, contextless text.
That’s just one failure mode. you may have seen some of these:
- sentences != facts. wording choice matters. it’s information compression. subtle inconsistencies enter and grow over time.
- projects change shape quickly, abstractly, and without explicit transitions.
- memory systems can’t deal with this kind of fuzzy, so initial brainstorming overstays its welcome
- you typically can’t inspect or clean it up yourself
- mem systems crave sequence, so any ingestion without correct timestamps will confuse it
- poor or no ability to handle “cognitive dissonance” - multiple perspectives on reality. In text memory, everything is or isn’t.
- topic grouping or tagging is inherently lossy and gets stuck like cement in the shape at initial ingestion
no doubt the engineers building memory system are talented. the work is clever.
but it’s not the ONE COOL TRICK that solves context engineering.
watch out for anything that promises the properties of ML training without the mechanism of ML training.
know how it works, so you don’t invest in a big KB only to find it’s making things worse.