Auto-memory
Build-along · ~8 min at your own keyboard
The model itself remembers nothing between sessions. It is stateless: each new session starts from zero. The model is the brain that writes the words; the harness is the program around it (Claude Code) that reads files, runs tools and decides what the model sees. Memory is a harness feature: Claude Code keeps a MEMORY.md and reads it back in at the start of each session. Say 'remember that…' and it attempts to save the fact. Check the file to see what it actually wrote, because a fact that isn't in the file doesn't exist next session. The discipline is choosing what belongs in durable memory versus what's just today's context.
The long version: Auto-memory in Claude Code ↗
Do this
0/4Checkpoints
- ▸You should now see Claude confirm it saved your fact.
- ▸You should now see a fresh session recall the fact without being told.
- ▸You should now see the memory file with only durable facts left after pruning.
The three ways this goes wrong
- The fresh session does not remember. — Say 'remember that ...' explicitly and check it confirms. Then fully exit and restart claude in the same project folder.
- Memory fills up with today's tasks. — Open the memory file and delete anything session-specific. Keep only facts that stay true.
- You cannot find the memory file. — Ask Claude: 'where is your memory stored for this project?' It will give you the path.
Where these ideas come from: Stateless, Model, Harness (AI Hero dictionary). Paraphrased here.
Share what happened
Paste what you asked it to remember, the fresh session's recall, and one thing you pruned.
Strip API keys, passwords, and tokens before you paste. Swap them for [KEY].
🔒 Don't paste confidential, personal-health, or payment info.