Hybrid retrieval, link graphs, typed frontmatter, relations, and computed fields
in one Rust CLI—with an optional desktop workspace.
15 open-source Agent Skills for Claude Code, Codex, Gemini CLI, Cursor, and GitHub Copilot.
A filesystem-native retrieval and data layer. Keep human-readable files as the authority; add fast search, graph context, typed queries, and computed frontmatter around them.
Markdown files are rows. Frontmatter is typed data. Links and relation fields are graph edges. Portable derived indexes live beside the collection.
One CLI, no database server and no container stack. HNSW vectors, BM25 segments, schema, and derived graph state stay local.
Structured JSON, a Rust library, watch mode, incremental ingest, graph expansion, and project-local skills fit directly into agent loops.
Frontmatter becomes typed columns. Markdown links become relations. Formula, Lookup, and Rollup fields are computed safely and materialized back into Markdown, so every query sees the same source-of-truth values.
---
title: August platform work
status: open
client: clients/acme.md
price: 240
quantity: 3
total: 720
---
# August platform workFROM invoices mdvdb collection invoices --recursive WHERE status = open --filter status=open ORDER BY total DESC --sort total --order desc LIMIT 20 OFFSET 0 --limit 20 --offset 0 --json This is a SQL-like workflow, not a SQL language: equality filters combine with AND, relations behave like foreign keys, and computed values can be filtered or sorted after they are materialized.
Follow a frontmatter link and populate the target like a one-hop JOIN.
Derive a typed value from fields in the same document.
Copy a field from one or more related Markdown documents.
Aggregate related values with an expression for counts, sums, averages, min/max, or another result.
Retrieval, memory, and structured operations all share the files your team and agents can already read and write.
Turn handbooks, product docs, runbooks, and decisions into section-level context. Hybrid search catches meaning and exact names; graph expansion follows citations; JSON output drops into any LLM tool loop.
$ mdvdb search "why did auth change?" --boost-links --expand 2 --json
Read the playbook →
Store daily logs and curated memories as ordinary Markdown. Ingest a single file after every write, filter by memory type, and apply exponential decay so old activity fades while maintained knowledge stays prominent.
$ mdvdb search "deployment preference" --filter type=memory --decay --decay-half-life 30
Read the playbook →
Model projects, research items, customers, content, or decisions with typed frontmatter. Relations connect records; formulas, lookups, and rollups compute state; collection queries filter, sort, and page the result.
$ mdvdb collection projects --recursive --filter status=active --sort risk --order desc --json
Read the playbook →
Hybrid RRF, semantic HNSW, lexical BM25, and semantic edge search. Results point to the matching heading, not merely the file.
A frontmatter .md path works like a foreign key and a graph edge. Populate the target and reverse references inline when you need JOIN-like context.
Time-based exponential decay for search scores. Recent files rank higher — ideal for AI agent memory where recency matters. Configurable half-life, path-based include/exclude.
Formula, Lookup, and Rollup modules atomically materialize typed values into Markdown. Filters, sorting, exports, and later ingests see the same values.
Discover stable automatic communities or define multi-label Topics with descriptions, seeds, thresholds, and an explicit Unassigned bucket.
Name a recursive folder scope without creating another index. Each Shard gets strict local graph analysis, local communities, and independent Topics.
OpenAI, OpenRouter, Gemini, Azure OpenAI, Bedrock, Hugging Face, Ollama, or a custom endpoint—with live model discovery where supported and automatic dimensions.
Watch changes, preview ingest, stream NDJSON progress, estimate reindex cost, inspect schema and sync health, and diagnose broken relations.
Semantic features use your chosen embedding provider. Lexical search, metadata queries, and many inspection commands remain available without a query-time embedding call.
Retrieve beyond isolated chunks. Follow authored links, frontmatter relations, semantic edge context, and multi-hop neighborhoods across the Collection or inside a Shard.
Discover stable communities automatically, then layer on your own multi-label Topics and thresholds.
Body links and frontmatter foreign keys share one graph, so backlinks, orphans, boost, and expansion agree.
Analyze the whole Collection or a named Shard with its own strict topology, communities, and Topics.
mdvdb stays CLI-first and library-friendly. Tesseract adds a visual workspace on top of the same files and the same CLI contract—without turning your vault into a proprietary database.
Search, query frontmatter, manage Shards and Topics, run computed modules, inspect providers, export graphs, and automate every result with JSON.
mdvdb is a derived layer around durable Markdown—not a new place you have to migrate your knowledge into.
Open it in a text editor, review it in Git, sync it however you like, and keep using existing Markdown tools.
.md + YAML frontmatter + links Vectors, BM25 segments, schemas, clusters, Topics, and graph state can be rebuilt from the collection.
.markdownvdb/ Human-readable output for terminals and stable JSON for scripts, model tools, desktop clients, and the Rust API.
CLI + JSON + Rust library Use Tesseract for editing, tables, graphs, Topics, Shards, and visual workflows without making it a dependency.
Optional Tesseract app | Source of truth | Your Markdown and declared schema overlay |
|---|---|
| Computed writes | Declared, module-owned, hash-guarded, and atomic |
| Secrets | Environment or .env files—never the YAML config |
| Scope | Folders or named Shards over one shared Collection index |
| Portability | No hosted database or proprietary document format required |