How my agents remember things
Five memory tiers with no extra infrastructure, and seven memory systems I turned down.
The usual advice is to add a vector database. For a software factory that's usually wrong. Most memory failures happen when writing. If the agent never wrote the decision down, no embedding will find it. So I asked for the least infrastructure that lets the factory remember what matters.
The answer is five tiers. Tap each one.
The five tiers I run
Every tier is a file, a git repo or a built-in text index, with no extra service or LLM bill.
No vector search
To recall past work, my agents run full-text search over every session transcript, using the index SQLite ships for free. It answers in tens of milliseconds and costs nothing. When you look for "that time I fixed the auth redirect", you remember a word that was in it. Semantic search helps only when the query and the memory share meaning but no words. That's a specific failure you can name. Until it shows up, full-text search is enough.
Global memory holds invariants only. Everything else goes in git.
The always-loaded memory file has a hard cap of a couple thousand characters, about six facts. The cap forces a question on every write: is this a permanent rule about how the factory works, or project knowledge? It's almost always project knowledge, so it goes in a decisions log in that project's git history, where it can be diffed, attributed and reverted. The cap keeps global memory from becoming a junk drawer.
What I turned down, and what would change my mind
Several memory systems are good, and I expect to adopt Supermemory eventually. For now I've rejected each one against a written trigger that would make it worth adding. Tap each system to see why I passed and what would change my answer.
Turned down for now, with a revisit trigger
Each one brings a daemon, a database, an embedding key or a background bill, and a problem you can name has to justify that.
The rule
Add an external memory provider only when you can name a recall that failed: the agent needed a fact it had stored and couldn't find it. "It might scale better" doesn't count. Until then, each extra system is one more thing to run and back up, with a bill that grows whether it helps or not.
The rest of the factory works the same way: each new part needs evidence, and the build stays cheap to throw away. A forgotten vector store is the classic line item someone questions a year later.