Nate B. Jones's open-source Open Brain (OB1): a Supabase Postgres database with a thoughts table — raw content, a 1536-dim pgvector embedding, JSONB metadata, and a content fingerprint for dedup — fronted by an open MCP server so any AI client (Claude Desktop, ChatGPT, Claude Code, Cursor, Codex) can read and write the same memory. Jones's own framing is blunt about what it isn't: "This isn't a notes app. It's a database with vector search and an open protocol." On capture, an LLM classifier — the "bouncer" — extracts metadata in parallel with the embedding, so the store gets sorted the moment a thought lands rather than degrading into an unsearchable dump.
Radical simplicity and radical cost. Our research clocks the whole build at roughly thirty to forty-five minutes, no coding required, on a stack that costs about five dollars to stand up — a free Supabase tier plus a five-dollar OpenRouter credit for embeddings and metadata extraction. It ships a real, working import catalog most systems only plan: recipes for ChatGPT export, Gmail, Obsidian vaults, X/Twitter, Instagram, Google Takeout, and Perplexity history, plus Slack and Discord capture bots for always-on quick capture. And because the protocol layer is MCP rather than a proprietary API, the same row-and-vector store is genuinely provider-neutral — Claude, GPT, and Gemini clients all read and write one brain without re-plumbing.
Mimir's spine is a direct descendant of OB1's: a row carrying raw content plus governed metadata, classified the instant it's captured rather than filed later by hand, with a dedup contract guarding against the same thought landing twice. Both reject "which folder?" in favor of "throw it in, the classifier sorts it" — OB1's bouncer and Mimir's deterministic-first facet inheritance are two different mechanisms aimed at the identical failure mode, an unsearchable landfill. Both are provider-neutral by design: OB1 through open MCP, Mimir through its own agent-facing serving core. And OB1's recipe catalog is the direct ancestor of the import-recipe pattern our research surfaces across the field — Mimir's own "federated sources" roadmap item is reaching for exactly what OB1 already ships.
Everything past the row. OB1's `thoughts` table is a single flat shape with one access tier — RLS-locked, service-role only, no classification ladder — where Mimir stamps four facets on every row (tags, domain, classification, principal) and enforces a public/internal/confidential/private ladder at the database level. OB1 has no concept of a standard being force-injected into an answer; Mimir's compartments and compulsory lenses exist specifically to make governance un-skippable rather than trusting anyone to remember it (see the governance walkthrough). OB1 stores rows, not a graph — no typed entities, no edges, no multi-hop reasoning; Mimir's Constellation is thin today but it is a real graph. And OB1's metadata is LLM-classified per atom, which our research flags as the exact mechanism that invites taxonomy drift at real scale — one run tagging client/SteadyStars and another tagging client/ss — where Mimir's deterministic-first inheritance is built specifically to avoid that failure mode.
Cost and speed, without qualification. OB1 is a working system for about five dollars and under an hour of setup; Mimir is a Postgres-plus-serving-daemon build that took a great deal longer to reach a single seeded compartment. OB1's semantic retrieval is live, cosine-similarity search over real embeddings; Mimir's pipeline, schema, HNSW vector index and search function are all built and working, but the embedding column sits at 0% coverage because the provider decision (hosted API vs. local) is still open, not because of any engineering gap — so Mimir is lexical-only until that closes, and filling it in afterward is a batch job over rows that already exist. OB1's import recipes are shipped and used; Mimir's federated sources are still a roadmap line. And OB1 has a real open-source community around it — a public repo, a Discord, contributed recipes and skills — where Mimir today is a single-operator build with no public repository at all. The honest summary: OB1 is proof that the bones of this idea run for the cost of a coffee subscription; Mimir is a bet that the bones aren't enough once the material gets sensitive enough to need a governance layer standing over them.
Grounded in our research on the active-curation school and Nate B. Jones's Open Brain project.