Nexus

Persistent memory and semantic search for AI coding agents. Local-first (no API keys required), three storage tiers with different lifetimes, an event-sourced document catalog with typed links between artefacts, and a specification-before-code decision-tracking workflow (RDR). Knowledge compounds across sessions instead of evaporating with the conversation.

The four-paragraph version is on the Tensegrity blog: How I actually use Nexus (concepts), Installing Nexus (ten-minute walkthrough), Nexus by Example (the pieces in practice).

At a glance

	What you get
Storage	T1 ephemeral session scratch (in-memory ChromaDB) · T2 SQLite + FTS5 for memory / plans / taxonomy / telemetry · T3 ChromaDB persistent + Voyage AI in cloud mode, ONNX MiniLM in local mode
Indexing	Tree-sitter AST chunking (23 languages) · CCE prose chunking (voyage-context-3) · PDF auto-routing (Docling → MinerU → PyMuPDF) · git frecency scoring · automatic topic discovery via HDBSCAN
Search	Semantic · keyword · hybrid · taxonomy-boosted · catalog-aware · plan-driven via nx_answer. Metadata filtering via --where bib_year>=2024 --where chunk_type=table_page and similar
Catalog	Event-sourced (events.jsonl canonical, SQLite is a deterministic projection) · typed links (cites, implements, supersedes, …) · Tumbler addressing inspired by Ted Nelson's Xanadu
Decision tracking (RDR)	Specification-before-code lifecycle: create → research → gate → accept → close, with post-mortems and a searchable history that surfaces during new design work
Claude Code plugin	13 specialized agents · 43 skills (RDR lifecycle, plan-centric retrieval, dev workflow) · 36 MCP tools across two servers · session hooks for cold-start context
CLI	nx with 16 top-level verbs (index, search, query, store, memory, scratch, catalog, taxonomy, enrich, doctor, upgrade, …)
Local-first	Default install runs entirely on your machine, zero API keys (ONNX MiniLM + local ChromaDB). Voyage AI + ChromaDB Cloud are opt-in for higher-quality embeddings
License	AGPL-3.0-or-later

By example

# Semantic search across indexed code, docs, and knowledge
nx search "how does authentication work"           # matches by meaning, not strings
nx search "error handling" --hybrid                # semantic + ripgrep, frecency-weighted

# Three tiers, one tool
nx scratch put "the bug is in the retry logic"           # T1, inter-agent session
nx memory put --project myapp --title "DB choice" "Postgres for concurrency"
nx store put --collection knowledge__myapp "Rate limit: 10k/min per vendor docs"

# Indexed PDFs filtered by bibliographic metadata
nx enrich bib knowledge__papers --delay 0.5
nx search "consensus protocols" --where bib_year>=2024

# Topic discovery is automatic; review keeps humans in the loop
nx index repo .
nx taxonomy review                                  # accept, rename, merge, split

Quick Start

Requires Python 3.12–3.13 and uv.

uv tool install conexus          # install the nx CLI
nx doctor                        # verify installation
nx index repo .                  # index your repo + discover topics (no API keys needed)
nx search "what does X do"       # semantic search with topic grouping, fully local
nx taxonomy status               # see discovered topics

Update: uv tool update conexus

Works immediately with local ONNX embeddings: no accounts, no API keys. For higher-quality cloud embeddings (Voyage AI), see the cloud setup instructions.

For Claude Code, install the plugin:

/plugin marketplace add Hellblazer/nexus
/plugin install nx@nexus-plugins

See Getting Started for the full walkthrough.

Three tiers, one lifecycle

Different information has different lifetimes. Together the three tiers form an integrated memory system that extends agent context across sessions and projects.

Tier	Purpose	Storage	API keys?
Scratch (T1)	Inter-agent session context: coordination and knowledge sharing across agent invocations	In-memory ChromaDB	No
Memory (T2)	Project-level persistence with full-text search	Local SQLite + FTS5	No
Knowledge (T3)	Permanent semantic knowledge: code, papers, docs, decisions searchable by meaning	Local ChromaDB (default) or ChromaDB Cloud + Voyage AI	No (local) / Yes (cloud)

Agents use all three tiers cooperatively. T1 enables inter-agent communication: sharing findings and preventing duplicate work within a session. T2 provides project decisions that constrain solutions. T3 surfaces how similar problems were resolved in other contexts. As the T3 knowledge base grows, it becomes the project's institutional memory, managing the information overload that accompanies complex designs and long-lived codebases.

What you can index

Code, documents, PDFs, and manual knowledge entries: anything that benefits from semantic search:

nx index repo .                          # code + docs + RDRs from a git repo
nx index pdf paper.pdf --collection knowledge__ml  # reference papers
nx store put --collection knowledge__ops "Redis maxmemory-policy: allkeys-lru for cache, noeviction for queues"

Repository indexing (nx index repo) is the most automated path. It classifies git-tracked files, chunks code into logical pieces via tree-sitter AST parsing across 31 languages, and embeds each chunk using local ONNX models by default, or Voyage AI models in cloud mode. Recently-touched files rank higher via git frecency scoring.

PDF indexing auto-detects math-heavy papers and routes them through MinerU (default-installed) for LaTeX extraction of equations; non-math PDFs use Docling directly. MinerU pulls ~2-3 GB of formula and table models on first use. To opt out of formula-aware extraction (e.g. for documents you know contain no math), pass --extractor docling explicitly. See PDF Extraction Backends for details.

See Repo Indexing for details and .nexus.yml configuration.

RDR: Research-Design-Review

Complex features are bigger than what fits in working memory, whether yours or an LLM's. Without a specification to track against, purpose drift sets in during implementation. An RDR is a specification document written before coding: it captures the problem, the research journey, competing options, and the chosen approach. Once accepted, the RDR is the stable target that implementation builds against. If the design proves wrong, abandon the RDR and draft a new one with what you learned; iteration lives in the chain of RDRs, not in any single document.

Each research finding is tagged with its evidence quality (verified against source code, supported by documentation only, or assumed) so readers know which conclusions are load-bearing and which need validation. The growing corpus is searchable, so prior decisions surface automatically when starting new design work.

RDR is fully optional. See RDR Overview for the full process.

Claude Code plugin

The nx/ directory is a Claude Code plugin that gives agents access to everything above. Install via the marketplace:

/plugin marketplace add Hellblazer/nexus
/plugin install nx@nexus-plugins

The plugin provides 13 specialized agents, 43 skills covering the RDR lifecycle, plan-centric retrieval, and development workflows, session hooks for automatic context initialization, and 36 MCP tools split across two focused servers: nexus (26 tools: search, store, memory, scratch, plans, 5 operator tools for structured extract/rank/compare/summarize/generate, plus 4 orchestration tools including nx_answer for plan-matched multi-step retrieval) and nexus-catalog (10 catalog tools: search, show, link, resolve, stats, etc.). The plan-centric retrieval layer (nx_answer) matches questions against a library of scenario templates, executes the matched plan, and records every run, so the library compounds with use. Agents search indexed code before proposing changes, check prior RDR decisions before designing new features, and coordinate through standard pipelines (plan → implement → review → test) with built-in quality gates.

The plugin integrates with Beads for task tracking. See nx/README.md for the full plugin documentation.

CLI Reference

The nx command provides direct access to all storage tiers, indexing, and search.

Command	What it does
nx search	Semantic and hybrid search across indexed code, docs, and knowledge. Supports --where with =, >=, <=, >, <, != operators for metadata filtering
nx index	Index git repos, PDFs, and markdown into searchable collections
nx enrich bib	Backfill bibliographic metadata (year, venue, authors, citations) from Semantic Scholar
nx enrich aspects	Extract structured paper aspects (problem, method, datasets, baselines, results) into T2
nx store	Store, retrieve, export, and import knowledge entries
nx memory	Per-project persistent notes (local, no API keys)
nx scratch	Inter-agent session context (in-memory, no API keys)
nx collection	Inspect and manage cloud collections
nx config	Credentials and settings
nx upgrade	Run pending database migrations and T3 upgrade steps
nx doctor	Health check: verifies dependencies, credentials, connectivity, schema
nx hooks	Install git hooks for automatic re-indexing on commit
nx catalog	Document registry: search, link, audit the catalog metadata layer
nx taxonomy	Topic taxonomy: discover, project across collections, review, merge, split
nx context	Project context cache: topic map for agent cold-start acceleration
nx mineru	MinerU server lifecycle (start/stop/status) for PDF extraction

Full details: CLI Reference.

Documentation

Document	What it covers
Getting Started	Install, local usage, Claude Code plugin, semantic search setup
CLI Reference	Every command, every flag
Storage Tiers	T1/T2/T3 architecture and data flow
Memory and Tasks	T2 memory, beads integration, session context
Repo Indexing	File classification, chunking pipeline, frecency scoring
Configuration	Config hierarchy, .nexus.yml, tuning parameters
Architecture	Module map, design decisions
Contributing	Dev setup, testing, code style
How I actually use Nexus	Blog overview: the substrate as a control surface for developers, teams, and agents
Installing Nexus	Blog install walkthrough: prerequisites, CLI, plugin, short tour
Nexus by Example	Blog walkthrough of the steam-driven semantic engine in practice

RDR (Research-Design-Review):

1. Overview: What RDRs are, when to write one, evidence classification
2. Workflow: Create → Research → Gate → Accept → Close
3. Nexus Integration: How storage tiers and agents amplify RDRs
4. Templates: Minimal and full examples, post-mortem template
5. Project RDR Index: All project RDRs with status

Prerequisites

• Python 3.12–3.13, uv, git
• For the Claude Code plugin: Node.js (provides npx) for the bundled sequential-thinking and context7 MCP servers, which are spawned via npx. The nx CLI alone does not need it.
• For cloud embeddings (optional): ChromaDB Cloud + Voyage AI accounts (free tiers available)
• For hybrid search: ripgrep

License

AGPL-3.0-or-later. See LICENSE.