Building a persistent memory system for Claude out of plain markdown files sounds like a weekend project. After six months of daily use, the real engineering challenge has become clear: it's not how you store memories, it's what happens when they contradict each other.
The approach itself is straightforward. Instead of relying on Claude's built-in context (which resets between sessions), you maintain a folder of .md files - each one storing facts, preferences, or project context. At the start of a session, you load the relevant files into the conversation. Claude reads them and behaves as if it remembers your history.
Why Conflicting Facts Break the System
Storage and retrieval are solved problems. A simple folder structure with descriptive filenames works fine. The genuine difficulty shows up when your stored facts contradict each other.
Say you noted in February that a client's budget was $5,000 per month. In April you updated it to $8,000. Now you have two files making opposite claims. Claude doesn't know which one to trust. Depending on which file loads first or which gets more weight in the context window - the pool of text the model can process at once - it might confidently act on the outdated number.
Human memory handles this naturally. We know newer facts supersede old ones, and we feel uncertainty when we're unsure. AI memory systems don't have that intuition unless you build it in explicitly.
The practical fixes aren't elegant. Timestamp every fact entry. Use a single canonical file per topic rather than appending new files. Build a regular review pass to manually reconcile contradictions before they compound. Some practitioners add a "supersedes" field to each entry, so the system can flag conflicts rather than silently picking one.
What This Means If You Use Claude for Ongoing Work
Claude Code ships with a built-in markdown memory system in its project settings - the same basic pattern, now formalized. Anyone using Claude for client work, research, or writing runs into this problem eventually. The first month feels like the system works great. By month three, stale or contradictory context starts producing confident but wrong outputs.
The core lesson from six months of real use: treat memory files like a database, not a diary. Append-only logs create conflicts. Canonical records with explicit update discipline stay trustworthy. The difference between the two approaches is small upfront and large over time.