Agent Memory
10 articles in series
Ten essays on persistent memory for AI agents — core vs archival memory, graph-native stores like Cognee and Zep/Graphiti vs plain vector search, what an agent should forget, the privacy question nobody answers, and a minimal memory architecture for your first agent.
A Memory Architecture for a Solo Builder's First Agent
You don't need a graph database on day one. A concrete, minimal memory stack for shipping your first agent without over-building or under-building it.
Memory as a Product Moat: Why Sticky Agents Remember You
A model can be copied in a weekend. A year of accumulated, structured memory about a user cannot. That asymmetry is where real agent moats live.
Building Your Own Memory Layer vs Buying One
Postgres plus embeddings can get a small team surprisingly far. Here's where that stops working and a dedicated framework starts paying for itself.
Temporal Graphs: Teaching an Agent That Facts Change Over Time
Zep and its Graphiti engine treat memory as a timeline, not a snapshot — so an agent can know a fact was once true without confusing it for current.
The Privacy Question Nobody Answers When They Ship Agent Memory
Persistent agent memory means storing personal facts indefinitely. Most teams ship it before deciding what consent, deletion, and ownership actually mean.
Session Memory vs Long-Term Memory: Picking the Right Layer
Not every agent needs a permanent memory store. A practical framework for deciding when ephemeral context is enough and when you need durable memory.
What an Agent Should Forget (and Why That's the Harder Problem)
Persisting memory is the easy half. Deciding what to decay, contradict, or delete is where most agent-memory systems quietly fall apart.
Why Graph-Native Memory Is Winning Over Pure Vector Stores
Embedding similarity finds facts that sound alike. It doesn't know they're connected. Graph-native memory layers like Cognee are built for the difference.
Core Memory vs Archival Memory: Letta's Two-Tier Model
Letta splits agent memory into always-in-prompt core blocks and queryable archival storage. It's a budget allocation problem disguised as an architecture choice.
Memory Is the New UX Pattern for AI Products
Model quality has converged across vendors. What users actually feel now is whether a product remembers them — and that's a UX problem, not a benchmark.