v0.13.0 — Remove vault_capture

Breaking change

Removed vault_capture MCP tool and neurostack capture CLI command.

The tool wrote directly to {vault_root}/inbox/ without git tracking, indexing, or any path back to origin. In runtimes where the vault is a read-replica (e.g. an indexer container pulling from git), captured files were stranded on the replica's filesystem and invisible to every other tool — including NeuroStack's own search and ask. The architectural promise that your Markdown files are never modified had a single, surprising exception; this release restores it.

Migration

Drop captured thoughts into your vault using your editor, your shell, or any tool that produces Markdown. The inbox/ folder convention is unchanged — only the built-in writer is removed.

Removed surfaces

• vault_capture(content, tags) MCP tool
• neurostack capture <thought> CLI subcommand
• capture_thought() and _make_slug() library functions (src/neurostack/capture.py deleted)
• CloudClient.vault_capture()
• vault_capture dispatch entry and manifest tool entry
• tests/test_capture.py

Tool count

NeuroStack now exposes 20 MCP tools (was 21).

---

Full changelog: https://github.com/raphasouthall/neurostack/blob/main/CHANGELOG.md

v0.11.1 — Critical Bugfix Release

Critical fixes from 10-agent stochastic audit

Security

• OAuth PKCE verification — /oauth/token never verified code_verifier against the stored code_challenge, making PKCE completely broken. Auth codes are now opaque, PKCE-bound, one-time-use tokens with 10-minute TTL. Redirect URI validation added against registered clients.

Billing

• MCP rate limiting tier resolution — All MCP tool calls hardcoded tier: "free", causing pro/team users to be incorrectly throttled. Tier is now resolved from Firestore for all 9 rate-limited tools.

Data Integrity

• Cloud sync _merge_memories data loss — INSERT OR REPLACE silently deleted embeddings, revision counts, and merge history on UUID conflict. Replaced with INSERT ... ON CONFLICT DO UPDATE SET to preserve existing fields. Fixed ttl_hours → expires_at column mismatch. Switched from raw sqlite3.connect() to get_db() for proper WAL mode and busy timeout.

Developer Experience

• embed_url default port — Default was 11435 (off-by-one from Ollama's standard 11434), causing embeddings to silently fail. Fixed in config, install scripts, and docs.
• vault-template not packaged in wheel — neurostack demo, neurostack init, and all profession packs were broken on pip installs. Templates now ship inside the package with a fallback to repo root for dev checkouts.

Housekeeping

• Synced server.json and manifest.json versions to 0.11.1

Full Changelog: v0.11.0...v0.11.1

v0.11.0

Full Changelog: v0.10.4...v0.11.0

v0.10.4 — Phase 1 Cloud Sync Foundation

Phase 1: Cloud Sync Foundation

Client-side infrastructure for the sync pipeline.

Bug Fixes

• Fixed ghost entries — push() now transmits diff.removed file list to the server, so deleted vault files are purged from the cloud index

New Features

• CloudClient 21-tool parity — all MCP tools now work in cloud mode via REST dispatch
• Cloud dispatch layer — automatic mode-based routing (local SQLite vs cloud API)
• Init default flipped to cloud mode — new users default to cloud; existing users unaffected
• API contract v2 spec — full specification for 9 sync endpoints (upload, sync, query, delete, download, memories, artifacts, push-lock, db-version)

Stats

• 9 new tests (299 total, 0 regressions)
• ~1,800 new LOC across 9 files
• Requires neurostack-cloud >=0.9.1 for full sync features (backwards-compatible with v0.9.0)

v0.10.3 — Windows support

New Features

• Windows support — NeuroStack can now be installed and run on Windows via npm install -g neurostack

Installation layer (npm)

• Added win32 to supported platforms
• uv auto-install uses .zip download + PowerShell Expand-Archive on Windows
• Fallback installer uses PowerShell (irm | iex) instead of sh
• command -v replaced with where for command lookup
• CLI wrappers created as .cmd batch files instead of bash scripts
• Platform-aware PATH separator, config directories, and setup instructions

Python runtime

• Config and data directories use AppData\Local\neurostack on Windows
• CLI wrapper creation produces .cmd files on Windows
• VS Code session harvesting uses %APPDATA%\Roaming\Code\User on Windows

Bug Fixes

• vault_search depth="full" validation error fixed (from v0.10.2)

Full Changelog

v0.10.2...v0.10.3

v0.10.2 — Fix vault_search full-depth validation

Bug Fixes

• fix(mcp): wrap vault_search full-depth return in dict — The depth="full" code path returned a list instead of a dict, causing Pydantic output validation to fail. This was the last remaining case missed by the list-to-dict fix in v0.10.1 (commit 0d6a751). All other depth modes (auto, triples, summaries) were unaffected.

Full MCP Tool Audit

All 21 MCP tools tested end-to-end. 20/21 passed — the only failure was the bug fixed above.

v0.10.1 — Unified init command

What's New

One-command setup

neurostack init now handles the entire onboarding flow:

npm install -g neurostack
neurostack init

The interactive wizard walks through:

1. Cloud or Local? — cloud uses Gemini, local uses Ollama
2. Lite or Full? — lite = FTS5 only, full = embeddings + summaries + communities
3. Ollama models — choose embedding and LLM models, auto-pull
4. Vault path + profession pack — point to existing notes or start fresh
5. LLM endpoint config — works with any OpenAI-compatible provider
6. Full index pipeline — automatically runs summaries, triples, and community detection

Non-interactive mode for scripted use:

neurostack init --mode full --pull-models ~/brain
neurostack init --cloud ~/brain

neurostack install deprecated

Still works but prints a notice pointing to neurostack init.

Bug fixes from v0.10.0

• Fixed Pydantic validation errors in vault_triples, vault_memories, and vault_related MCP tools (wrapped list returns in dict)

Full Changelog: v0.10.0...v0.10.1

v0.10.0 — Neuroscience-grounded ranking & community detection

What's New

Neuroscience-grounded reranking (replaces cross-encoder)

Three new scoring signals in hybrid_search, using only existing embeddings in SQLite — zero external models, zero network calls:

• Energy landscape convergence confidence — measures how representative the matched chunk is of its note's overall embedding distribution. Based on Hopfield attractor basin dynamics: deep narrow wells (low chunk variance) score higher than broad shallow ones.
• Lateral inhibition — winner-take-all diversity penalty. Higher-ranked results suppress semantically similar competitors, promoting result diversity. Mirrors PV+/SOM+ inhibitory interneuron function during engram allocation.
• Prediction error demotion — notes with unresolved prediction errors (previously observed poor semantic fit) are demoted. Closes the predictive-coding feedback loop: errors logged at retrieval time now influence future ranking.

Attractor basin community detection (replaces Leiden)

Hopfield-style attractor dynamics replace the Leiden algorithm for community detection:

• Blended similarity matrix: S = 0.6·cosine + 0.25·co-occurrence + 0.15·wikilinks — richer signal than Leiden's entity co-occurrence alone
• Inverse temperature β controls granularity: β=0.5 for coarse themes, β=2.0 for fine sub-themes
• Lateral inhibition merges singleton communities into nearest neighbours

Dependencies removed

Dependency	License	Why removed
sentence-transformers	MS MARCO (non-commercial)	Replaced by native scoring signals
leidenalg	GPL-3.0	Replaced by attractor basin clustering
python-igraph	GPL-2.0+	No longer needed

The [community] optional dependency tier is eliminated — communities now work with just [full] (numpy only).

Install mode simplification

neurostack install now offers two modes instead of three:

• Lite — FTS5 search + graph, no ML
• Full — + embeddings, summaries, communities

Full Changelog: v0.9.5...v0.10.0

v0.9.5

Full Changelog: v0.9.4...v0.9.5

v0.9.4

Full Changelog: v0.9.2...v0.9.4