where_defined
Jump straight to where any symbol is declared — across the whole repo.
Claude forgets you after every conversation. NeuroVault doesn't.
100% local. Your memory is embedded on your own machine and never leaves it — no cloud, no API keys, no account. Private by architecture, not a setting you switch on.
A local-first memory layer for AI agents. Notes, decisions, people, projects — captured once, recalled whenever any MCP-speaking agent asks. Analytics mode shows the structure of your brain at a glance: bigger nodes for what gets referenced most, soft tints for clusters of linked notes.
macOS says "NeuroVault is damaged"? It's the missing Apple
certificate (we're open-source), not malware — drag the app to
Applications, then run
xattr -dr com.apple.quarantine /Applications/NeuroVault.app
once, or right-click → Open.
Details ↗
Every memory is a plain .md file on your disk. The database is
just an index. If the index ever breaks, we rebuild it from the files.
You own your brain.
Every note, decision, person, and project becomes a node. Each one wires to what it relates to — until the shape of how you think is something you can actually see.
Every new session, your agent forgets. So you compensate — and the workarounds cost you more context than they save.
CLAUDE.md that bloats every prompt — and still goes stale.Graphify parses your repo on-device with tree-sitter and builds a living graph of every symbol, import, and call. Now your agent reasons about your code's structure — not just greps its text. No source ever leaves your machine.
where_defined
Jump straight to where any symbol is declared — across the whole repo.
who_calls
Every caller of a function or method, so the agent traces real usage.
blast_radius
Change this and what breaks? The transitive set of everything that depends on it.
whats_in_file
The symbols and signatures in any file — a structured table of contents.
fuse
Link a decision or wiki note to the code it governs — memory meets source.
RustPythonTypeScriptTSXGoJavaC#Ruby
LongMemEval is brutal — long multi-session histories, facts that get updated and contradicted, questions that need temporal reasoning. Across the full 470-question set, NeuroVault surfaces the right memory in its top 5 results 93.8% of the time, and in the top 10 98.1% of the time — running entirely on your machine. Every number reproduces on your laptop with one command.
Measured on LongMemEval-S with NeuroVault's real recall() path —
the exact code your agent calls — using 100% on-device embeddings. The
stricter recall@5 (every gold passage, not just one) is published too:
0.861. The full harness and a per-question receipt for
every answer live in the repo.
Full methodology, harness & per-question receipts ↗
No cloud, no account, no telemetry — just plain markdown files on your disk. The backend binds to loopback only and refuses outside connections.
Every note is a node in a live, force-directed graph. Analytics mode sizes nodes by influence (PageRank) and tints clusters of linked ideas (Louvain), so the shape of your knowledge is visible at a glance.
Hybrid search fuses semantic, keyword, and graph hits, reranks them, and weights by recency — so recall surfaces what you meant, not just what you literally typed.
Live preview, [[wikilinks]], a ⌘K command
palette, VS Code-style tabs, and 7 themes. It behaves like the editor you
already use all day.
Claude drafts wiki pages from your raw notes and names your clusters, running in the Claude / Cursor session you already have open via MCP. No API keys. Every change is a diff you approve.
One click writes NeuroVault's MCP entry into ~/.claude.json
(your login untouched). The agent auto-starts the backend on first use —
no "is the server running?". Just recall().
Work, research, personal — switch brains with one keystroke,
Obsidian-compatible. Opt a project folder into its own brain with a
.neurovault file; each codebase gets memory scoped to itself.
A tiny floating control parks in the corner: backend status, start / pause, open the app. Shrink it to just the logo, or hide it entirely.
Anything you'd keep in notes, NeuroVault makes askable — and because it's all on your machine, even the private, proprietary, and privileged stuff is fair game.
Thousands of notes, but search can't find what you phrased differently years ago. Ask in plain language and get the actual passage back — even when no keyword matches.
Point NeuroVault at any folder of markdown and it builds a searchable brain in minutes. Recall finds the right note by what you meant, not the words you typed.
Run graphify on any repo for an on-device code knowledge graph, then ask where_defined, who_calls, or blast_radius before you refactor. Source never leaves the machine.
Build a linked web of literature notes across years of work. The graph and semantic recall surface every note on a theme across chapters — and reveal where two sources disagree.
Ask your diary when your anxious entries cluster, or what you said last time you felt burned out. The embeddings and retrieval run entirely on-device — nothing reaches a cloud LLM.
Save your preferences, decisions, and context once as markdown you own. Every session your agent recalls it over MCP — across Claude, Cursor, and Codex, not locked to one vendor.
Even Obsidian-plus-AI ships your note text to a cloud LLM. NeuroVault keeps the files and the search local — so lawyers, accountants, and therapists can ask about privileged notes with nothing leaving the device.
Save the repo layout, auth flow, and why-we-did-X once; the next session recalls them instead of inventing wrong file paths. Per-project brains keep one repo's context out of another's.
And it scales with you. Add as many notes as you want — your vault is plain markdown (10,000 notes is only ~130 MB), the database is a lightweight rebuildable cache over it, and recall stays in the low‑hundreds of milliseconds even on a big brain.
NeuroVault speaks MCP, so Claude Code, Claude Desktop, Cursor, and Codex all read and write the same brain. Tools are grouped into tiers — pick a slice, not the kitchen sink. Switch any time in Settings.
recall · related · session_start
+ remember · status · switch_brain · update
+ chunk & temporal recall · duplicate check · core memory · history
power users+ graphify · maintenance · graph editing · contradictions · images
manual wikilinks, entity
co-mentions, and semantic embedding matches each have
their own colour. Hide them by default to read the structure;
bring them back when you want the full neighbourhood.
⌘K is the primary nav verb. Three sections in one
palette — Commands (fuzzy), Notes (title search),
Memory (semantic recall after three characters).
auto-approve toggle and let trusted compiles ship
straight to the vault.
Auto-saved as markdown. A file watcher triggers the ingestion pipeline — chunk, embed locally, extract entities, update the knowledge graph.
A UserPromptSubmit hook silently runs your message through an extractor.
Eight patterns catch preferences, decisions, deadlines, stacks.
Each fact becomes a first-class kind='insight' engram
with a wikilink back to where you said it.
Claude calls recall() via MCP. Hybrid search fuses
semantic, keyword, and graph hits. Cross-encoder reranks the top
candidates; PageRank importance lifts hub notes when Analytics mode
is on. Answers now carry context they couldn't have had before.
The same local-first engine, synced through a private Git repo you already own. Drop in your meeting notes and decisions; NeuroVault turns them into a knowledge graph every teammate's agent can recall — "what did we promise Acme, and when?" answered with sources.
127.0.0.1:8765. It refuses connections from other machines.