Ditching Vector DBs: Building Agent Memory with a Folder of Markdown

gimalay · reddit · 2026-08-02

The author proposes an alternative agent memory system to traditional vector databases: using a folder of Markdown files as a database. Standard stacks often require structured DBs, text files, and vector DBs, leading to complex glue code and hard-to-debug black-box retrieval.

The author argues that most agent memory needs are actually structured queries (e.g., "what tasks are still open"), which are essentially WHERE clauses rather than similarity searches. The core design includes:

File paths as primary keys and frontmatter as the schema, supporting Mongo-style operators.
Links as typed edges, where inline links are references and lone links imply containment, enabling complex traversals.
Agent-specific optimizations: --max-tokens caps context length, --expect 1 forces write operations to declare affected rows to prevent errors, and agents can query built-in docs dynamically.

The author honestly notes the limitations: it is not semantic search (though BM25 is built-in), but at personal scale (loading 20k docs in 0.7s), once the store has real structure, the need for semantic similarity is surprisingly rare.

Original post →

Ditching Vector DBs: Building Agent Memory with a Folder of Markdown

More from coding & agent