neurostack 0.3.0

Build, maintain, and search your knowledge vault. CLI + MCP server with stale note detection, semantic search, and neuroscience-grounded memory.

Looking for sponsors — NeuroStack is built and maintained by a single developer. If it's useful to you, consider sponsoring the project or buying me a coffee to help fund development, hosting, and new features.

Your second brain, starting today. Install one command, answer a few questions, and you have a knowledge vault that gets smarter every day you use it.

NeuroStack helps you start and grow a personal knowledge base — whether you have thousands of notes or none at all. It scaffolds a vault, indexes everything by meaning, surfaces what needs attention, and gives your AI tools long-term memory. It works with any AI provider — use it as an MCP server, as a CLI tool alongside Claude, or pipe its output into any LLM workflow. Everything runs locally. Your notes never leave your machine.

Get started

npm install -g neurostack
neurostack install
neurostack init

1. npm install -g neurostack — bootstraps the CLI (handles Python, uv, and all dependencies automatically).
2. neurostack install — interactive wizard to choose your installation mode and optionally set up Ollama models.
3. neurostack init — walks you through vault location, LLM config, and profession packs.

No prior config needed. No Python, git, or curl required.

Installation modes

Mode	What you get	Size	GPU needed?
lite (default)	FTS5 search, wiki-link graph, stale note detection, MCP server	~130 MB	No
full	+ semantic search, AI summaries, cross-encoder reranking	~560 MB	No (CPU inference)
community	+ GraphRAG topic clustering (Leiden algorithm)	~575 MB	No

The install command handles everything — dependency syncing via uv, Ollama installation, model pulls, and config updates. It detects what's already installed and skips unnecessary work:

# Interactive — walks you through mode, Ollama, and model selection
neurostack install

# Non-interactive — specify mode directly
neurostack install --mode full

# Full mode + pull Ollama models in one shot
neurostack install --mode full --pull-models

# Custom models
neurostack install --mode full --pull-models --embed-model nomic-embed-text --llm-model qwen3:8b

When choosing full or community mode:

• Ollama not installed? — offers to install it automatically (Linux) or shows the download link (macOS).
• Models already pulled? — skips the download and moves on.

You can re-run neurostack install at any time to upgrade between modes (e.g., lite → full).

Alternative install methods

# PyPI
pipx install neurostack              # isolated environment
pip install neurostack               # inside a venv
uv tool install neurostack           # uv users

# One-line script (lite mode)
curl -fsSL https://raw.githubusercontent.com/raphasouthall/neurostack/main/install.sh | bash

# One-line script (full mode)
curl -fsSL https://raw.githubusercontent.com/raphasouthall/neurostack/main/install.sh | NEUROSTACK_MODE=full bash

Note: On Ubuntu 23.04+, Debian 12+, and Fedora 38+, bare pip install outside a virtual environment is blocked by PEP 668. Use npm, pipx, uv tool install, or create a venv first.

To uninstall:

neurostack uninstall

What it does

Build — start your second brain in minutes

• Interactive setup — neurostack init walks you through vault location, model selection, and profession packs. No docs to read first.
• Scaffolds your vault — profession-ready templates and folder structures so you're not staring at an empty folder.
• Onboards existing notes — point it at any folder of Markdown files and it adds frontmatter, generates indexes, and sets up the vault structure. No manual migration.
• Works with what you have — connects to any Markdown vault. Use Obsidian, Logseq, or plain .md files.

Maintain — your vault gets better every day

• Finds stale notes — surfaces notes that keep appearing in searches where they don't belong. The signal to review before they mislead you or your AI.
• Remembers what matters — recent, active notes get priority. Old, unused notes fade — just like real memory.
• Maps your knowledge — reveals hidden connections and topic clusters you never planned.
• Watches for changes — auto-indexes new and updated notes so your vault is always current.

Search — find anything by meaning, not just keywords

• Searches by meaning — finds notes by what they say, not just what they're titled. Ask a question, get an answer.
• Works with any AI tool — use as an MCP server with Claude Code, Codex, Gemini CLI, Cursor, Windsurf, or any MCP-compatible client. Or pipe CLI output into any LLM workflow.
• Saves tokens — tiered retrieval sends your AI key facts first, full notes only when needed. 96% fewer tokens per query than naive RAG.

How it works

NeuroStack indexes your vault into a knowledge graph, then uses techniques borrowed from memory neuroscience to help you maintain it:

What you see	What it does	How your brain does it
Stale note detection	Flags notes that appear in the wrong search contexts	Prediction error signals trigger memory reconsolidation
Hot notes + decay	Recently active notes get priority; unused notes fade	CREB-elevated neurons preferentially join new memories; excitability decays over time
Memory dedup	Detects near-duplicate memories, merges on demand	Pattern separation in dentate gyrus distinguishes similar but distinct memories
Topic clusters	Reveals thematic groups across your vault	Neural ensembles form overlapping assemblies
Smart retrieval	Starts with key facts, escalates only when needed	Hippocampal rapid binding complements slow cortical learning
Context recovery	Rebuilds task-specific context after interruption	Hippocampal replay reinstates activity patterns during memory retrieval
Meaning-based search	Finds notes by concept, not just keywords	Associative memory retrieval follows semantic paths

See the neuroscience citations

Feature	Key Paper
Stale note detection (prediction errors)	Sinclair & Bhatt 2022, PNAS 119(31)
Hot notes (excitability windows)	Han et al. 2007, Science 316(5823)
Knowledge graph (engram networks)	Josselyn & Tonegawa 2020, Science 367(6473)
Community detection (neural ensembles)	Cai et al. 2016, Nature 534(7605)
Tiered retrieval (complementary learning)	McClelland et al. 1995, Psychological Review

Full citations: docs/neuroscience-appendix.md

Profession packs

NeuroStack ships with profession-specific starter packs — domain templates, seed notes, and workflow guidance so you're not starting from a blank vault.

# The interactive setup offers profession packs automatically
neurostack init

# Or apply to an existing vault
neurostack scaffold researcher

# See what's available
neurostack scaffold --list

Each pack adds:

• Templates — domain-specific note formats (e.g., experiment logs, synthesis notes)
• Seed research notes — interconnected examples showing the vault in action
• Extra directories — folder structure tailored to the workflow
• CLAUDE.md guidance — workflow instructions for your AI assistant

Pack	Description	Templates	Seed Notes
researcher	Academic or independent researcher — literature reviews, experiments, thesis work	synthesis-note, experiment-log, method-note, paper-project	6
developer	Software developer or engineer — architecture decisions, code reviews, debugging	architecture-decision, code-review-note, debugging-log, technical-spec	6
writer	Writer or content creator — fiction, articles, worldbuilding, craft notes	article-draft, character-profile, story-outline, world-building-note	5
student	Student or lifelong learner — lectures, study guides, courses, exam prep	assignment-tracker, course-overview, lecture-note, study-guide	5
devops	DevOps engineer or SRE — runbooks, incidents, infrastructure, change management	change-request, incident-report, infrastructure-note, runbook	6
data-scientist	Data scientist or ML engineer — analyses, models, datasets, experiment tracking	analysis-note, dataset-note, model-card, pipeline-note	6

Contributions of new profession packs are welcome. See CONTRIBUTING.md.

Token economy

NeuroStack's tiered retrieval sends your AI only what it needs:

Depth	Tokens per result	When it's used
Triples	~15 tokens	Quick factual lookups — 80% of queries resolve here
Summaries	~75 tokens	When you need more context
Full content	~300 tokens	Deep dives into specific notes

Compared to naive RAG (dumping full document chunks at ~750 tokens each), NeuroStack uses 96% fewer tokens per query. That means lower API costs, faster responses, and more of your AI's attention on actually answering your question instead of processing background noise.

Agent memories

Your AI can write back short-lived memories - observations, decisions, conventions, bugs - that surface automatically in future vault_search results. Unlike vault notes, memories are lightweight and can expire.

MCP tools: vault_remember, vault_forget, vault_memories, vault_update_memory, vault_merge

# CLI
neurostack memories add "deployment requires VPN" --type convention
neurostack memories add "auth token expires in 1h" --type observation --ttl 7d
neurostack memories search "deployment"
neurostack memories list
neurostack memories update <id> --content "new text" --add-tags "infra"
neurostack memories merge <target_id> <source_id>
neurostack memories forget <id>
neurostack memories prune              # Remove expired memories
neurostack memories stats

Entity types: observation, decision, convention, learning, context, bug

Memories with a --ttl auto-expire after the given duration. Without TTL, they persist until explicitly forgotten or pruned.

Update in place

Edit a memory without deleting and recreating it. Update any field - content, tags, entity type, workspace, or TTL. Content changes automatically re-embed for semantic search. Tag operations support replace, add, and remove.

neurostack memories update 42 --content "revised observation"
neurostack memories update 42 --add-tags "auth,security"
neurostack memories update 42 --remove-tags "old-tag"
neurostack memories update 42 --type decision --ttl 0    # make permanent

MCP tool: vault_update_memory - pass only the fields you want to change.

Dedup and merge

When saving a memory, NeuroStack checks for near-duplicates using a two-stage pipeline (FTS5 keyword overlap, then cosine similarity on embeddings). It never auto-merges - instead it returns candidates so the caller can decide.

# Save returns near-duplicate warnings
neurostack memories add "use env vars for secrets"
#   Saved memory #5 (observation)
#   ! Near-duplicates found:
#     #2 (similarity: 0.76)
#       Always use environment variables for secrets
#   Merge: neurostack memories merge <target> <source>

# Explicit merge - folds source into target
neurostack memories merge 2 5

Merge unions tags, keeps the longer content, picks the more specific entity type, and tracks an audit trail (merge_count, merged_from).

MCP tools: vault_remember (returns near_duplicates in response), vault_merge

Tag suggestions

When saving a memory, NeuroStack suggests tags based on FTS5 overlap with existing tagged memories and file path extraction. No LLM call - pure heuristic, under 10ms.

neurostack memories add "Fixed handler in src/auth/middleware.py" --type bug
#   Saved memory #6 (bug)
#   Suggested tags: py, auth, bug, security
#   Apply: neurostack memories update 6 --add-tags py,auth,bug,security

MCP tool: vault_remember returns suggested_tags in response.

Session harvest

Extract insights from your Claude Code session transcripts automatically. NeuroStack scans the JSONL transcripts for decisions, bugs, conventions, and learnings, deduplicates against existing memories, and saves them.

neurostack harvest                  # Harvest from the latest session
neurostack harvest --sessions 5     # Scan the last 5 sessions
neurostack harvest --dry-run        # Preview what would be saved
neurostack harvest --json           # Machine-readable output

MCP tool: vault_harvest - call from your AI assistant to extract and save insights from the current session.

Automation hooks

Schedule periodic harvest runs with a systemd user timer instead of manually running neurostack harvest:

neurostack hooks install            # Install hourly harvest timer
neurostack hooks status             # Check if timer is active
neurostack hooks remove             # Remove the timer

Context recovery

After /clear or starting a new conversation, use vault_context to recover task-specific context. Unlike session_brief (a time-anchored status snapshot), vault_context is task-anchored - it retrieves memories, triples, summaries, and session history relevant to a specific task, within a token budget.

neurostack context "implement auth middleware" --budget 2000
neurostack context "database migration" --workspace work/my-project --no-triples

MCP tool: vault_context - returns structured JSON with memories, triples, note summaries, and session history scoped to the task.

Excitability decay

Notes that haven't been accessed recently lose their "hotness" score through exponential decay - just like biological memory. The decay command reports which notes are dormant, active, or never used, using the existing hotness scoring system as the single source of truth.

neurostack decay                              # Default report
neurostack decay --threshold 0.1 --half-life 60  # Custom thresholds
neurostack decay --json                       # Machine-readable

vault_stats also includes excitability breakdown (active/dormant/never-used counts).

Features at a glance

Feature	Lite (no GPU)	Full (local AI)
Full-text search	Yes	Yes
Wiki-link graph + PageRank	Yes	Yes
Stale note detection	Yes	Yes
Session transcript search	Yes	Yes
Session insight harvesting	Yes	Yes
Semantic search (by meaning)	—	Yes
AI-generated summaries & triples	—	Yes
Cross-encoder reranking	—	Yes
Topic clustering (Leiden)	—	+community

What gets installed

npm install -g neurostack bootstraps the CLI, then neurostack install lets you choose your mode. Here's exactly what ends up on your machine:

Lite mode (default)

Installed with neurostack install --mode lite (or the default interactive selection).

Component	Location	Size	What it is
npm wrapper	$(npm root -g)/neurostack/	~20 KB	Node.js bin + install scripts
uv	~/.local/bin/uv	~30 MB	Python package manager (auto-installed)
Python 3.12	~/.local/share/uv/python/	~50 MB	Standalone Python managed by uv (not your system Python)
NeuroStack source	~/.local/share/neurostack/repo/	~2 MB	Python source + vault templates
Python venv + deps	~/.local/share/neurostack/repo/.venv/	~50 MB	pyyaml, watchdog, mcp, httpx
Config	~/.config/neurostack/config.toml	~200 B	Vault path, Ollama endpoints, model name
Database	~/.local/share/neurostack/neurostack.db	Grows with vault	SQLite + FTS5 knowledge graph

Total: ~130 MB. No GPU needed. Keyword search, wiki-link graph, stale note detection, and MCP server all work in lite mode.

Full mode

Installed with neurostack install --mode full. Adds ML dependencies on top of lite.

Component	Location	Size	What it is
Everything from lite	(see above)	~130 MB
numpy	in .venv/	~30 MB	Numerical computing
sentence-transformers	in .venv/	~100 MB	Embedding and reranking models
PyTorch (CPU)	in .venv/	~300 MB	ML runtime (CPU-only, pulled by sentence-transformers)
Ollama (external)	User-installed	~1 GB+	Required for embeddings and summaries — install separately

Total: ~560 MB + Ollama. Adds semantic search, AI-generated summaries/triples, and cross-encoder reranking.

Community mode

Installed with neurostack install --mode community. Adds community detection on top of full.

Component	Location	Size	What it is
Everything from full	(see above)	~560 MB
leidenalg	in .venv/	~5 MB	Leiden community detection (GPL-3.0)
python-igraph	in .venv/	~10 MB	Graph analysis library (GPL-2.0+)

Total: ~575 MB + Ollama. Adds topic cluster detection across your vault.

Clean removal

neurostack uninstall    # Removes source, venv, database — preserves config

Config at ~/.config/neurostack/ is kept so a reinstall picks up where you left off. To remove everything: rm -rf ~/.config/neurostack.

CLI

NeuroStack is a command-line tool. Every feature is available from your terminal:

neurostack install                  # Install/upgrade mode and Ollama models
neurostack install --mode full --pull-models  # Non-interactive install
neurostack init                     # Interactive setup wizard
neurostack init [path] -p researcher  # Non-interactive with profession pack
neurostack onboard ~/my-notes       # Onboard an existing folder of notes
neurostack onboard ~/my-notes -n    # Dry run — preview changes first
neurostack scaffold researcher      # Apply a profession pack to existing vault
neurostack scaffold --list          # List available profession packs
neurostack index                    # Build the knowledge graph
neurostack search "query"           # Search by meaning or keywords
neurostack search -w "work/" "query"  # Search scoped to a workspace path
neurostack harvest                   # Extract insights from recent sessions
neurostack memories list              # List agent write-back memories
neurostack memories update <id> --content "new text"  # Update in place
neurostack memories merge <target> <source>  # Merge two memories
neurostack context "task description"  # Task-scoped context recovery
neurostack decay                     # Excitability report - dormant vs active notes
neurostack graph "note.md"          # See a note's connections
neurostack prediction-errors        # Find stale or misleading notes
neurostack communities query "topic"  # Explore topic clusters
neurostack brief                    # Morning briefing - what needs attention
neurostack record-usage note.md     # Mark notes as used (drives hotness scoring)
neurostack hooks install            # Set up automatic harvest timer
neurostack stats                    # Index health overview (includes excitability)
neurostack doctor                   # Validate all subsystems
neurostack watch                    # Watch vault for changes, re-index automatically
neurostack serve                    # Start as MCP server for AI tools
neurostack update                   # Pull latest source + re-sync dependencies
neurostack uninstall                # Complete removal — data, deps, and npm package

More commands

neurostack tiered "query"           # Tiered search: triples → summaries → full
neurostack triples "query"          # Search structured knowledge triples
neurostack summary "note.md"        # Get a note's AI-generated summary
neurostack communities build        # Run Leiden detection + generate summaries
neurostack communities list         # List detected topic clusters
neurostack backfill [summaries|triples|all]  # Fill gaps in AI-generated data
neurostack reembed-chunks           # Re-embed all chunks
neurostack folder-summaries         # Build folder-level context summaries
neurostack sessions search "query"  # Search session transcripts
neurostack harvest                             # Extract session insights as memories
neurostack harvest --sessions 5 --dry-run      # Preview without saving
neurostack memories add "text" --type observation  # Store a memory
neurostack memories search "query"  # Search memories
neurostack memories list            # List all memories
neurostack memories update <id> --content "revised"  # Update in place
neurostack memories update <id> --add-tags "new-tag"  # Add tags
neurostack memories merge <target> <source>  # Merge two memories
neurostack memories forget <id>     # Remove a memory
neurostack memories prune           # Remove expired memories
neurostack memories stats           # Memory store overview
neurostack context "task" --budget 2000  # Task-scoped context recovery
neurostack decay                    # Excitability report
neurostack decay --threshold 0.1    # Custom dormancy threshold
neurostack hooks install            # Set up hourly harvest timer
neurostack hooks status             # Check automation status
neurostack hooks remove             # Remove automation hooks
neurostack demo                     # Interactive demo with sample vault
neurostack status                   # Overview of your vault and config

Use with any AI provider

NeuroStack is provider-agnostic. Your vault is a local SQLite database with a CLI and MCP interface — use it however fits your workflow.

MCP server (Claude Code, Codex, Gemini CLI, Cursor, Windsurf, etc.)

Add to your MCP config and your AI assistant gets long-term memory from your vault. Works with any MCP-compatible client:

• Claude Code MCP setup
• Codex MCP setup
• Gemini CLI MCP setup

{
  "mcpServers": {
    "neurostack": {
      "command": "neurostack",
      "args": ["serve"],
      "env": {}
    }
  }
}

CLI (works with everything)

The CLI outputs plain text — pipe it into any AI tool or workflow:

# Use with Claude Code as a CLI tool
# See: https://docs.anthropic.com/en/docs/claude-code/cli-usage
neurostack search "deployment checklist"

# JSON output for scripting — all query commands support --json
neurostack --json search "query" | jq '.[] | .title'

# Scope to a workspace path (or set NEUROSTACK_WORKSPACE env var)
neurostack search -w "work/" "deployment"
export NEUROSTACK_WORKSPACE=work/my-project

# Pipe into any LLM
neurostack search "project architecture" | llm "summarize these notes"

# Use in scripts, CI, or automation
CONTEXT=$(neurostack tiered "auth flow" --top-k 3)
echo "$CONTEXT" | your-preferred-ai-tool

All 16 MCP tools

Tool	What it does
vault_search	Search your vault by meaning or keywords, with tiered depth
vault_summary	Get a pre-computed summary of any note
vault_graph	See a note's neighborhood - what links to it and what it links to
vault_triples	Get structured facts (who/what/how) extracted from your notes
vault_communities	Answer big-picture questions across topic clusters
vault_context	Assemble task-scoped context for session recovery
vault_stats	Check the health of your index (includes excitability + memory stats)
vault_record_usage	Track which notes are "hot" (recently accessed)
vault_prediction_errors	Surface notes that need review
vault_remember	Store a memory (returns near-duplicate warnings + tag suggestions)
vault_update_memory	Update a memory in place (partial updates, re-embeds on content change)
vault_merge	Merge two memories (unions tags, tracks audit trail)
vault_forget	Remove a memory by ID
vault_memories	List or search stored memories
vault_harvest	Extract insights from Claude Code session transcripts
session_brief	Get a compact briefing when starting a new session

Configuration

Config lives at ~/.config/neurostack/config.toml:

vault_root = "~/brain"
embed_url = "http://localhost:11435"
llm_url = "http://localhost:11434"
llm_model = "qwen2.5:3b"

Every setting has a NEUROSTACK_* env var override.

Architecture

~/your-vault/                        # Your Markdown files (never modified)
~/.config/neurostack/config.toml     # Configuration
~/.local/share/neurostack/
    neurostack.db                    # SQLite + FTS5 knowledge graph
    sessions.db                      # Session transcript index

NeuroStack never modifies your vault files. All data (indexes, memories, sessions) lives in its own databases.

Requirements

• Linux or macOS
• npm install: Just Node.js — everything else is installed automatically
• Full mode: Ollama with nomic-embed-text and a summary model (e.g., phi3.5)

Links

• Website: neurostack.sh
• Contact: [e-mail]

License

Apache-2.0 — see LICENSE.

Community detection extra

The optional neurostack[community] extra installs leidenalg (GPL-3.0) and python-igraph (GPL-2.0+). These are not installed by default and are isolated behind a runtime import guard.

If you install neurostack[community], you are responsible for complying with GPL terms when redistributing. The core NeuroStack package remains Apache-2.0 and contains no GPL code.

Contributing

See CONTRIBUTING.md.