MemoFS Stores AI Agent Memory as Local Files to Fix Session Amnesia

AI agents routinely lose context between sessions because large language models rely on token-based working memory rather than persistent storage, causing them to forget past decisions and architectural rules. Expanding context windows to over one million tokens has not solved the problem, instead introducing issues like context compaction, attention degradation, and rising API costs. Vector databases, a common alternative, are difficult to inspect, lack version control integration, and perform poorly on exact-match queries involving file paths or symbol names. MemoFS addresses these limitations by storing all agent memory as plain Markdown and JSON files inside a project's local .memofs/ directory, making memory human-readable, grep-able, and version-controllable. The tool uses a hybrid BM25 and vector index alongside a knowledge graph to give agents durable, auditable, and branch-aware memory without relying on remote databases.