Why do AI agents still feel like they have the memory of a goldfish?
One minute they're refactoring complex logic, the next they've completely lost the context of what they were doing.
We kept running into this while working with AI agents in real-world environments, so we decided to break it down from a systems perspective—not just prompts, but what's actually happening under the hood.
Here’s the video if you want the full walkthrough.
The real problem isn’t prompts
A lot of devs try to fix this by writing longer prompts or adding more context.
That helps… until it doesn't.
The real issue usually comes down to how memory is handled:
- What gets persisted
- What gets cached
- What gets discarded between interactions
If you don't design for that, your agent will always feel inconsistent.
Two different approaches: Claude Code vs OpenClaw
We looked at two very different philosophies:
Claude Code
- Structured, layered memory system
- Relies on a 4-level architecture
- More predictable, but also more constrained
OpenClaw
- Agent-driven memory (more autonomous)
- Decides what to store and reuse
- Feels more flexible, but comes with trade-offs
Neither is “better” universally, it depends on how much control vs autonomy you want.
Memory vs Caching
(this is where things get interesting)
One thing that gets mixed up a lot:
- Memory → long-term persistence across sessions
- Caching → short-term reuse for efficiency
Most production issues we’ve seen come from confusing these two.
The part people are not talking about: Security
As soon as you give agents memory, you're also increasing risk.
One example we cover is indirect prompt injection:
- Malicious data gets stored as “memory”
- The agent trusts it later
- Behavior gets manipulated without obvious signals
This becomes especially important if your agent touches production systems.
What we’ve learned building with this
A few takeaways that have held up for us:
- Memory needs boundaries, not just storage
- Caching should be intentional, not automatic
- More autonomy = more responsibility (and more risk)
- Observability is not optional if you're going to production
We're curious how others here are approaching this.
Are you comfortable giving AI agents persistent memory in production yet?
Top comments (0)