Proposed 'memoryfield' file format stores agent memories as Markdown pages

The introduction of memoryfield changes how agent memory is handled by treating it as a portable file format rather than a multi-stage pipeline. Engineers can now store memories as ordinary files that fit into existing version control and backup workflows. This moves the responsibility of memory management from complex services to the file system.

Each memoryfield page is written directly by the agent in Markdown, optionally enriched with YAML frontmatter for metadata. An accompanying SQLite file can hold vector embeddings for semantic lookup, removing the need for separate pgvector or Neo4j instances. The design limits each page to roughly 8 KB, which matches typical vector embedding sizes and encourages granular, topic-focused pages.

Existing approaches suffer from three main flaws: they tie memory to a specific vendor harness, they rely on heavyweight stacks that are hard to administer and scale, or they distill facts into context-free graphs that lose usefulness. Memoryfield sidesteps these issues by avoiding vendor lock-in, minimizing external dependencies, and preserving the original prose context. This makes the format easier to audit, version, and migrate across different agent implementations.

Adopting memoryfield requires agents capable of writing coherent Markdown and a workflow that can handle many small files; the per-page size ceiling means large documents must be split, which may increase file count. Engineers must also decide whether to include the SQLite index for search, adding a lightweight dependency but keeping the overall setup simpler than current alternatives. The format’s reliance on plain text also means that any corruption or loss of a file directly affects the associated memory, underscoring the need for robust backup practices.