Here’s a problem that drove me quietly mad. I use several different AI coding tools, switching between them depending on the task. And every single one starts each session with complete amnesia. The tool that helped me fix a bug on Tuesday has no idea what the other one designed on Monday. For weeks, the only shared memory in the system was me — re-explaining context over and over. I was the bottleneck.
So I built what I call a memory bridge. At its heart it’s simple: a folder of structured notes — one per working session — that every AI tool can read from and write to. I work across several machines — a couple of laptops and a couple of rented cloud servers — and the memory bridge is what ties them together. Save a session on my laptop in the morning, and a tool running on a server in the afternoon can recall what happened.
I’m not an engineer, so I made choices that favored “boring and reliable” over “clever”:
I skipped the fashionable approach. The standard way to make searchable memory is embeddings and a vector database. I used neither. My search is plain keyword matching. It’s boring, it works offline, and — the part I care about — it can’t invent a match that isn’t actually in the files. When an agent recalls something, that thing definitely exists.
I separated saving from indexing. My first version rebuilt the master index every single time anything was saved. It was slow and created a mess. The fix was to make saving fast and append-only, and rebuild the index separately, on a schedule, in a batch. Clean and quick.
Short takeaways beat full transcripts. Each saved session gets a few lines of distilled “here’s what we learned” — five to seven short bullets — that show up in the master index. The AI tools scan those first and only open the full file when it’s actually relevant. This matters more than it sounds: AI context is limited and expensive, so you don’t spend it reading everything to find the one thing you need.
One hub, not fifteen installs. Early on, every tool on every machine — laptops and cloud servers alike — needed its own copy of everything. Centralizing into one hub that they all connect to collapsed that into a single place to maintain.
The system now holds hundreds of sessions, and it quietly changed how I work. I can start something with one AI tool in the morning and pick it up with a different one in the afternoon, and the second one already knows what happened. The tools stopped feeling like separate gadgets and started feeling like a team that shares a notebook.