← Writing
November 1, 2026·3 min read

Memory You Never Read Is Just Storage

Three sessions in one day started with the word 'continue', and each time my AI assistant had no idea what we'd been doing. The memory system was fine. Nothing ever read from it.

Today I typed "continue" into three separate Claude Code sessions. Each time the assistant told me there was nothing earlier in the session to continue from, and offered a menu of unrelated projects pulled from a memory index. The thread I actually meant was small: getting the TrueNAS box to come back on its own after a power loss.

The first two sessions made it worse by being helpful. One told me the problem was blocked on a JetKVM ATX power board I didn't have. The next one said memory showed the problem resolved in August: I own the DC extension instead, and its "Restore Power Loss: Power ON" setting already repowers the box when AC returns. Both answers came from the same memory store. They disagreed because the assistant only looked at whichever notes happened to be loaded, and never searched for the latest one.

So I stopped the TrueNAS work and asked a different question: why does this keep happening when I've built a memory system for exactly this?

The audit was short. The palace held about 3,500 stored items and passed its integrity check. Hooks saved transcript content when a session ended or compacted, and a nightly job synced notes in. Writing worked fine. But there was no read path. Nothing at session start pulled recent context, nothing in my config told the assistant to search before answering a vague prompt, and the diary, the one place that records where we left off, had 21 entries because only I ever wrote to it. The auto-saved content was raw transcript rather than summaries, so even a good search returned noise.

I'd built a very reliable filing cabinet and never told anyone to open it.

The fix had three parts. A rule in my global config: on a vague prompt like "continue", search the last three days and read the diary before answering, and do the same before talking about any past project or decision. A SessionStart hook that injects recent diary entries and newer items so the assistant begins with context instead of having to remember to ask for it. And a habit of writing a diary entry at every checkpoint (shipped, decided, blocked) with the next step and what it's blocked on, rather than once at the end of a session that might crash first.

I should be honest about how much of this is proven. The hook only takes effect at the next session start. Later the same night, in a session that had lost its context to a crash, I still had to say "you can look in mem palace for everything." And the hook is built to fail silently if the palace is unreachable, which is the right call for startup latency, but it means a down memory server looks identical to an empty one. I haven't solved that yet.

The TrueNAS question, for what it's worth, turned out to be mostly answered already. The MAC address was in the JetKVM console's IPv6 address, the Wake-on-LAN entry was already saved, and the BIOS setting needs a reboot I haven't scheduled.

The general lesson is that retrieval is a separate system from storage and it needs its own owner. Teams make this mistake with documentation, and people make it with notes apps. If your agent, your wiki, or your team's runbook only has a write path, you don't have memory. You have a log, and the first person who needs it at 2am will reconstruct everything from scratch, confidently and slightly wrong.