Why AI Agents Lose Their Memory And How MemoFS Solves It

Short summary

AI agents suffer structural amnesia because context windows act as RAM, not durable storage, and context compaction silently destroys rationale and past decisions. MemoFS proposes storing agent memories as plain Markdown and JSON files in a .memofs/ directory, with derived full-text indexes, vector embeddings, and entity graphs. This approach enables git-native version control, inspectable memory via standard file tools, and hybrid BM25+vector search for exact symbol matching.

• •Context windows are working memory, not storage — compaction destroys subtle rules and past decisions
• •Vector databases are opaque, lack git integration, and fail at exact keyword/symbol matching
• •MemoFS stores memories as plain files in .memofs/ with derived indexes, enabling inspection, versioning, and rollback

Generated with AI, which can make mistakes.