4 Kinds of Agent Memory
Post by Paweł Huryn on LinkedIn — “My agent has 4 kinds of memory. 3 live in files.”
The 4 Memory Types
| Memory Type | What It Is | Where It Lives |
|---|---|---|
| Working Memory | What the agent holds right now — context window, goals, scratchpad | You manage it (context window) |
| Semantic Memory | Facts & knowledge — your strategy, rules, what “good” looks like | Markdown files, versioned in git |
| Episodic Memory | What happened before — past runs, decisions, outcomes | A table (CSV in git, not a database) |
| Procedural Memory | How to do things — reusable steps loaded on a trigger | Skills — also files |
“Three of those are just files in git: two folders of markdown, one table. The fourth is the context window. You probably don’t need a vector database either.”
Two More (Escape Hatches)
When there’s too much to read:
- Vector store — searches a million documents by similarity. Enterprise search, not for your agent.
- Graph database — relationships with labels (owns, depends on) at scale. But your linked markdown is already a graph the agent walks by reading.
Key Takeaway
“That’s my setup: markdown files plus one CSV. No vector DB, no embeddings, no server. It sharpens every week because the agent writes its own lessons back into the files.”
Most agent memory doesn’t need a vector DB — files in git are enough.