Nodedex

NodeDex

your project's decision history, for agents

Give your agent your project's decision history: it checks what was already ruled out before proposing, follows supersede edges to current truth instead of citing a stale decision, and inherits the whole investigation cold — session after session. Built automatically from the agent's own conversations into a local causal graph. Works with Claude Code, Cursor, Hermes, or an agent loop you wrote yourself. (A different job than your fact store — run both.)

Quick start · How it works · Connect your agent · Why NodeDex · Evidence · Docs

An agent walking its own memory: a dead-end → why it was abandoned → the causal chain. The structure is the recall.

Without NodeDex, session N+1 quietly re-derives what session N already settled — which approaches failed (and why), which decisions were replaced (and by what), which constraints still bind. NodeDex keeps that record in a local SQLite graph the agent navigates deliberately, session after session.

Status — early & solo-built. NodeDex is developed by one person and is at an early stage. The engine is tested end-to-end (1319 passing tests), but it hasn't been battle-tested across many machines and agents yet — expect rough edges, and please open an issue when you hit one. Feedback at this stage is hugely valuable.

Newest, and least battle-tested: the three-wire setup — your agent installs NodeDex into itself (reflex + capture + gate), and NodeDex verifies each wire by what actually happened rather than taking the agent's word. The gate (a pre-edit check, warn-only, fail-open) is brand new; it replaces a roadmap item that used to read "hook-enforced check, not just a nudge". Whether it changes agent behaviour on a long task is exactly what we're measuring next — we don't claim it yet.

On the roadmap: interop with Claude Code's native memory · a "what's new since last session" surface · broader agent-host support.

Get started

Three steps. The first sets up the server (once); the other two connect your agent.

# 1. Install + run the setup wizard — asks for a provider key, model, port, and database.
npx nodedex                 # or: npm i -g nodedex && nodedex onboard

# 2. Point your agent at the running server (one line; this is the only host-specific part).
#    Claude Code shown — Cursor, Codex, Copilot, Hermes and custom loops in "Connect your agent".
claude mcp add --transport http nodedex http://127.0.0.1:3001/mcp   # 3001 = the port you chose

3. Say Set up NodeDex to your agent. It wires itself in — CAPTURE (so the graph fills), REFLEX (so it remembers to check), GATE (so it checks at the moment it edits) — asking your permission for each. NodeDex verifies each one actually happened; nothing is taken on trust.

That's the whole path. Details: Setup (the wizard) · Connect your agent (every host) · Try it in 60 seconds (a demo graph, no key). Switching tools later moves no data — point the new agent at the same server and say Set up NodeDex again.

Why NodeDex

Forgetting announces itself — the agent asks again, you sigh, you re-explain. The failure that actually costs you re-work is silent: the agent re-proposes an approach it tried and abandoned three sessions ago, or plans against a decision that was replaced two weeks back. The stale decision looks exactly as authoritative as the current one, the ruled-out approach sounds fresh, and the hours get burned twice.

NodeDex fixes being confidently wrong about the past. That's a knowledge-status problem — ruled-out, superseded, still-open — and status is structure, not recall. So NodeDex keeps the project's reasoning residue as a graph the agent walks: decisions with their why and the alternatives that lost, dead-ends as a permanent, enumerable closed-door list the agent is taught to check first (a strong nudge, not a hard block), constraints that can't silently move, and the causal chains tying them together.

A new agent inheriting a project doesn't get a ranked pile of facts — it walks the investigation, and learns what not to repeat.

Different job. Most "agent memory" stores facts and preferences and retrieves the top-k at query time — perfect for "the user prefers TypeScript", and genuinely good at it. NodeDex captures the reasoning around those facts: what was tried, what was ruled out, what was decided and why.

They're complementary, and many setups run both. NodeDex doesn't replace your fact store.

When you don't need NodeDex

Honest scope: for short sessions, one-shot tasks, and fact-recall workflows, native memory is good enough — skip NodeDex, it would be overkill. It earns its keep on long-running, decision-dense, multi-session work where repeating a failure is expensive — and especially on autonomous runs, where no human is watching to catch the agent re-proposing a dead end.

Try it in 60 seconds — no key, no setup

A fresh graph is empty (the value accumulates from your real work), so NodeDex ships a sample project history you can explore immediately:

npx nodedex demo

That serves a small finished project — its decisions (with the why and the rejected alternatives), its dead-ends, its constraints, and one decision that got superseded — at http://127.0.0.1:3009/mcp. Point any MCP agent at it and ask:

1. "What did we try and abandon in the ratelimiter project, and why?"
2. "Why was token bucket chosen — what else was considered?"
3. "Is 'keep the counters in Redis' still the current decision?" — it isn't; watch the agent follow the supersede edge to current truth instead of quoting the stale one.
4. "What's still open or unverified?"

If those answers are useful from a 15-block toy, imagine them from six weeks of your own work. nodedex stop 3009 ends the demo; your real graph is never touched.

Want receipts before installing? Evidence it works — what the pipeline captures from real transcripts, graph health, and search that admits its limits.

What it is

A local knowledge graph of what the project established — every decision, dead-end, constraint, and the causal links between them — persisted outside the context window and read by the agent over MCP. Not a note-taking tool: nobody curates it; a background pipeline builds it from the work itself.

The agent is stateless by default. NodeDex gives it a durable record of the project's reasoning that survives every context reset — the LLM is just the reasoning engine that runs on top of it.

Key properties:

• Storage is local (~/.nodedex/*.db). Extraction is local or cloud, depending on the model you pick: a cloud provider (e.g. OpenRouter) receives conversation content for extraction; a local model keeps everything on the machine. The database is plaintext unless you set NODEDEX_DB_ENCRYPTION_KEY.
• Agent navigates deliberately — tree view first, search second
• Causal chains are first-class — blocks carry the links that produced them (standalone records are allowed)
• Dead ends are retained by design — extraction can still miss or misclassify one; every block carries its verbatim source excerpt so you can check it
• Async AI pipeline — turns raw exchanges into structured knowledge without agent overhead

How it works — three actors

  Autonomous agent (e.g. Hermes)
        │   reads ▲                         │ each finished turn (a COPY)
        │  (MCP)  │                         ▼
        │   ┌─────┴──────────────┐   ┌──────────────┐
        └──▶│  NodeDex server    │◀──│ capture tee  │  (out-of-path, fire-and-forget)
            │  • MCP read tools  │   └──────────────┘
            │  • REST API        │
            │  • AI pipeline ────┼──▶ writes blocks / chains / links (async)
            │  • self-maintenance│   (dedup · provenance · heal)
            └─────────┬──────────┘
                      ▼
              ~/.nodedex/<your>.db   (SQLite WAL — local, bitemporal)

• The agent navigates the graph with read-only MCP tools (workspace_get, workspace_search, workspace_filter, workspace_tree, workspace_stats, …). Knowledge-write tools are hidden by default; a read records usage metadata (access counts) on the block it opens — never content, links, or status. The one deliberate write on the default surface: workspace_task_update — the agent maintains its own task status (only the agent knows when its work is actually finished; everything else stays the pipeline's job).
• The pipeline (a server-side AI) compiles everything — facts, decisions, dead ends, insights, constraints, reasoning chains — async, from each captured turn. The agent never has to stop and save.
• Capture wires each finished turn to the server. The path depends on the host: Hermes are captured by reading their own state.db (they ignore a model proxy); an OpenAI-compatible host that honors a custom base URL can route its model through NodeDex's proxy; an agent whose loop you control uses a small out-of-path tee (workspace_install_capture). NodeDex is a passive MCP server — it can't see the agent's replies on its own, so without capture the graph stays empty (see Connect your agent).

SQLite WAL — a single local file, bitemporal relations (history preserved, never deleted).

Setup

One command (Node ≥ 18):

npx nodedex

First run launches the setup wizard; once configured, the same command just starts the server. Prefer a real install? npm i -g nodedex puts nodedex on your PATH. Everything — config, key, databases, logs — lives in ~/.nodedex/, and npx nodedex uninstall removes it all.

Or let your agent install it. If you already run an agent with shell access (Claude Code, OpenClaw, Hermes…), paste AGENT-INSTALL.md into it: the agent asks you three consent questions, runs one headless setup command, starts the server, and connects itself.

Native driver note: the SQLite driver ships prebuilt for common platforms. Only if npm has to compile it from source do you need a C/C++ toolchain (macOS: xcode-select --install · Debian: build-essential python3 · Windows: VS Build Tools → C++).

git clone https://github.com/NodeDex/NodeDex-v0.1.git
cd NodeDex-v0.1/server && npm install          # compiles/fetches the SQLite driver
cd ../tui && npm install
npm run dev                                    # first run launches the setup wizard

The wizard walks you through it: choose your model — OpenRouter (cloud, bring your key, with a free-but-trains-on-prompts warning where it applies) or Local / self-hosted (Ollama / LM Studio / vLLM — offline, no key, $0; the wizard scans and lists your local models so you just pick one) → pick a free port → create or name a database → it starts the server on localhost and, on the final screen, shows you your server URL (and an auth token if you chose a network/Docker bind) — save these; they're exactly what you hand your agent in the next step. On first run it also downloads the bundled local embedding model (one-time) so semantic search works offline with no extra key.

The extraction model must be smart AND have a large context window — the pipeline asks it to comprehend whole conversation arcs and emit strict, complete structured output, which is harder than chat. From real testing: a 12B local model comprehends fine but reliably fails the strict structured passes (drops mandatory items, runaway output) — the pipeline refuses to save the incomplete result, so you get safety but no memory. Working floor: a cheap cloud model (the recommended gemini-2.5-flash-lite, ~½¢ per extraction, 1M context) or a capable ~27–30B+ local model served with a ≥16k real context window (watch Ollama's default num_ctx — imports often run at 8k regardless of what the model supports). For local runs, also set NODEDEX_THINKING_BUDGET=off and a NODEDEX_MODEL_CAPS entry for your model.

The wizard sets up a same-machine server (localhost) — the default. Running your agent in Docker or on another machine is a later, advanced setup (bind 0.0.0.0 + a token); see docs/how-to/connect-mcp-over-http.md.

Embeddings default to a bundled local model (offline, free, no key). To use a hosted embedder instead, put EMBEDDING_PROVIDER=gemini (or openai) in ~/.nodedex/.env — it's a config value, not a shell command, so it's the same on every OS. All config lives in ~/.nodedex/ and your provider key never enters the repo. (On Windows, ~/.nodedex means C:\Users\<you>\.nodedex.)

Next: connect your agent to the running server.

Connect your agent

NodeDex is built for agentic agents doing long, multi-session work — anything that can read a file, write a file, and run a tool. Claude Code, Cursor, Codex, Copilot, an autonomous host like Hermes, or a loop you wrote yourself: the setup below is the same for all of them, because your agent does it, not you.

There are exactly two steps.

Step 1 — point your agent at the server (the only host-specific part)

The wizard already started the server and printed your URL (http://127.0.0.1:<port>/mcp). Register it with your host over MCP (Streamable-HTTP). One line, once:

Use 127.0.0.1, not localhost — the server binds IPv4, and on Windows localhost resolves to IPv6 ::1 first, so it silently fails to connect. Lost your URL? Run nodedex connect; ~/.nodedex/connect-snippets.md also holds ready-to-paste configs, regenerated at each launch.

Step 2 — tell your agent to wire itself in

Say this to it, once:

Set up NodeDex

That's the whole instruction. Your agent installs NodeDex into itself — it knows its own host far better than we do, so it finds its own files and its own seams. It will ask your permission before touching anything, and it will be nagged on every tool result until it's done, so it's hard to skip.

There are three wires, and they have different lifetimes — which is the entire point:

The gate warns; it never blocks, and it fails open — if NodeDex isn't running, it does nothing at all. It will never stand between you and your editor.

None of it is taken on trust. NodeDex verifies each wire by what actually happened — it reads the file back and looks for the marker, waits for a real turn to land in the graph, waits for a real check to reach the endpoint. An agent that says it did the setup and didn't will keep being nagged.

Because connecting the tools is demonstrably not enough. In our own testing, an agent with NodeDex connected read this project's recorded dead-ends at 12:17, wrote the code at 16:44, and shipped the exact bug those dead-ends warned about. It understood them perfectly — in isolation it uses them every time. It simply no longer had them: the instruction to look arrived once, at connect, and four hours and several compactions later it was gone.

A memory an agent doesn't consult at the moment of decision is not memory. The reflex puts the habit where the host re-reads it every turn; the gate puts the reminder at the edit itself. Capture makes sure there is something there to find.

If your agent can't write files

Then it can't install anything, and you wire it by hand — three HTTP calls, any language:

# 1. REFLEX — paste this into that agent's system prompt (it must be present on EVERY turn)
curl http://127.0.0.1:3001/api/agent-reflex?format=text

# 2. CAPTURE — POST each finished turn, fire-and-forget, out of your model's path
#    POST /api/reflect/trigger  { user_message, agent_response, reasoning }

# 3. GATE — call before a file edit; feed anything it returns back into context
curl http://127.0.0.1:3001/api/gate/check

On Node, workspace_install_capture hands your agent a ready-made adapter (a fire-and-forget tee — your model call is never touched or slowed) and workspace_install_gate hands it the gate script.

Some hosts persist their own transcripts and expose no post-turn seam (Claude Code, Hermes). For those, NodeDex ships a watcher instead — read-only, forward-only, enabled from the TUI — so capture works with no wiring at all. The wizard detects them and asks.

Full detail / other hosts. Two things make it work, and they're separate:

1. Give the agent the tools (read side). Register the running server with your agent host over MCP (Streamable-HTTP at /mcp). How depends on where your agent runs:

Same machine (default)

Your agent runs on this computer, so it can reach the server on loopback — and the wizard already started it there. Point your host at the /mcp URL (most MCP hosts take a small JSON config):

{
  "mcpServers": {
    "nodedex": { "url": "http://127.0.0.1:3001/mcp" }
  }
}

Use 127.0.0.1, not localhost — the server binds IPv4, and on Windows localhost resolves to IPv6 ::1 first, so localhost silently fails to connect while 127.0.0.1 works.

Or let the host spawn it over stdio (npm i -g nodedex first, so the command exists on PATH; from a clone use node /path/to/server/scripts/nodedex-entry.mjs instead):

{
  "mcpServers": {
    "nodedex": { "command": "nodedex-server" }
  }
}

CLI-style host (e.g. Hermes): hermes mcp add nodedex --url http://127.0.0.1:3001/mcp, then reload its MCP connections and start a new session.

Claude Code, Cursor, VS Code — the agent picks the config file, not the editor. The Claude Code extension uses .mcp.json even when it runs inside Cursor or VS Code; Cursor's and VS Code's own agents use their own files. Same /mcp URL, different location + key per agent:

Switching hosts (e.g. Hermes → Claude Code) moves no data. The graph is host-agnostic — every agent reads and feeds the same graph. To add or switch, just connect the new host (MCP above + its capture path below) and it immediately sees everything the previous one left. Run several at once if you like.

Agent in Docker / on another machine (advanced — not first-run)

A containerized or remote agent can't reach the host's loopback, so the server must bind 0.0.0.0 and use a token, and the agent connects via the host address (e.g. host.docker.internal). That's a deliberate, later setup — full steps (bind host, token, Docker networking) are in docs/how-to/connect-mcp-over-http.md.

The agent now sees the read tools and can traverse the graph. (The server delivers its own usage protocol via the MCP instructions field — no prompt changes needed from you.)

2. Turn on capture (write side) — required, or the graph stays empty. NodeDex is a passive MCP server: it sees tool calls, not your agent's replies. Something has to push each finished turn to it. Pick the path that fits your agent:

(a) Watchers — zero-setup capture for hosts that persist their own turns

The lowest-friction path: NodeDex reads the host's own conversation log (read-only, on your machine, forward-only — never past history) and needs nothing from the host itself. The onboarding wizard detects installed hosts and asks which to capture; toggles live in the TUI's settings view.

• Claude Code — tails your session transcripts (~/.claude/projects/…jsonl). Every session becomes its own arc in the graph; the projects row scopes which projects are captured (* = all).
• Hermes — reads Hermes's state.db (Hermes ignores a custom model base URL and never invokes its shell hooks, so this is the only path that works for it). The sources row is the privacy filter (default tui). Full walkthrough: docs/how-to/connect-hermes.md.

(b) OpenAI-compatible host that honors a custom base URL — route the model through NodeDex

The zero-deploy path for a host that does let you redirect its model endpoint (Hermes does not — use (a)). Point your agent's model base URL at NodeDex's proxy: it relays each call to your real provider unchanged (no added latency, streaming works) and captures the turn in passing. In your agent's model/provider settings, set:

• base URL: http://127.0.0.1:3001/api (use 127.0.0.1, not localhost, on Windows: a 0.0.0.0-bound server is IPv4-only and localhost resolves to IPv6 ::1 first; a remote/Docker agent uses the host address instead)
• API key: your usual provider key (e.g. OpenRouter sk-or-…) — forwarded untouched, never stored
• model: unchanged

No file to deploy, and no NodeDex token needed for this path (the proxy is exempt and uses your own provider key).

(c) Agent whose code/loop you control (Agent SDK, LangChain, your own loop) — the tee

Have the agent call workspace_install_capture once; it returns a tiny out-of-path tee to drop into your post-turn seam (it POSTs {user, response, reasoning} to the server). On a token-gated server, set NODEDEX_TOKEN where the tee runs so its POSTs authenticate.

Without one of these, the agent can read memory but the graph never grows.

What you hand your agent depends on the path. Reading is always the MCP …/mcp URL (plus a token only on a Docker/network server). Capturing differs by host:

Using Hermes? Full end-to-end walkthrough (MCP read + the state.db watcher + every gotcha): docs/how-to/connect-hermes.md.

Full per-host detail (HTTP transport, Docker / host.docker.internal, the capture adapter, env toggles): docs/how-to/connect-mcp-over-http.md and docs/how-to/deploying-nodedex.md.

Reconfigure / uninstall

# change your key / model / provider (merges into ~/.nodedex/config.json)
npx nodedex setup --provider openrouter --key sk-or-...   # validated before saving
npx nodedex onboard                                       # or re-run the full wizard

# remove all local data + config (~/.nodedex: config, API key, databases, logs)
npx nodedex uninstall      # asks for confirmation — this is destructive

Or do it inside the TUI: settings view → provider row switches cloud/local and picks a model (local models are auto-scanned). A model change within the same provider applies live; a provider/endpoint switch applies when the server relaunches (switch db, or restart the TUI).

Multiple keys + automatic failover. Real setups keep more than one key — different accounts, and one held as a fallback. nodedex config opens a dedicated keyring page: store several API keys, mark one active and one fallback, set a fallback model, and choose what happens when the active key bills out (fail over and keep going, or respect the spend-pause). The running server then automatically fails over to the fallback key when the active one is rejected (bad key) or runs out of credit — same model, no restart. Setting the active key swaps the live server in place; everything reflects in nodedex tui (both read the same ~/.nodedex/config.json). The keyring is also reachable from the TUI settings keyring row.

Keys are stored in plaintext in ~/.nodedex/config.json today (single-user, local-first) — treat that file as a secret; encrypt-at-rest is planned before a wider release.

Uninstall does not remove the package (npm rm -g nodedex for that) or the NodeDex entry in your agent host's MCP config (remove that on the host). To change your key / model / provider, use the TUI (nodedex tui → 3 health → Enter on provider), or re-run nodedex onboard — or headless: nodedex setup --provider openrouter --key … --model ….

Commands

Server (cd server):

The nodedex command (from npx nodedex / npm i -g nodedex; from a clone, cd server && npm link gives you the same command):

The TUI has three views — 1 memory (browse roots, walk stories, / search), 2 feed (memory forming live + what the agent read), 3 health (server/db switch, provider

• model picker, pipeline knobs, capture toggles, review queue). Everything is a row: ↑↓ move, enter acts.

The TUI is the normal way to run the server (it configures the key/model + picks a port and database, and relaunches it on a db switch). The server/ scripts above are the manual/advanced path.

Block types

The pipeline classifies and writes these from your turns; the agent reads them.

How the agent reads it

Navigate first, search second — the tree shows everything.

workspace_tree                     # root view — projects + counts
workspace_filter(concepts)         # cold start: concepts → relevant roots
workspace_get(label, "relations")  # a block + its causal chain(s)
workspace_search(query)            # fallback: semantic + keyword + concept, ranked by match
workspace_stats(agent_id)          # graph landscape + extraction freshness + pending flags

A block alone is a headline; its chain is the story — workspace_get returns the named causal chain(s) the block sits on, so one read gives the whole arc.

Search is honest about its limits: every hit shows which root it lives under and its match signals, superseded blocks are flagged with what replaced them, and when nothing really matches the results say so (weak matches only) instead of posing as an answer.

Evidence it works

We test NodeDex the way it's actually used: extract reasoning through the pipeline → check the graph's health → see whether the agent can traverse to what it needs — read from the live graph, not inferred from a score.

Feeding research-paper derivations through the pipeline, here's the reasoning residue it captured per paper, by block type (read directly from the graph):

The captured profile mirrors the source's reasoning style — a bound-derivation paper is try → reject → choose → constrain (dead-end / constraint-heavy: 53 blocks, incl. 2 dead-ends, 10 decisions, 5 constraints, 8 chains); a clean theorem proof is establish → conclude (fact / chain-heavy). On a representative paper the graph was healthy — 55 blocks, only 1 flagged for review, decisions wired to their justifying facts, zero islanded roots — and navigating it (tree → root → decision → chain) handed back the scoped result with its derivation, not a decontextualized fragment.

Not a RAG pass-rate. We deliberately don't report a one-shot benchmark score: a passive "retrieve-and-inject" call can't exercise traversal, so its number measures the agent's task skill, not the memory. Full write-up: docs/NODEDEX-MEMORY-MODEL.md.

Engine health: 1319/1319 server tests pass, with extraction → graph → retrieval validated end-to-end — including capture-loss (ack-safe watcher cursors), the MCP read surface itself (what the agent receives, provenance included), and the setup wires (a claimed-but-unwritten reflex is caught, not believed). Tests prove the mechanisms work; whether the memory makes your agent measurably better is the claim we're still validating with users.

We also ran NodeDex on itself: the last ~50 turns of the Claude Code session that built it, captured by the watcher, extracted into a 223-block graph — audited at 9/10 fidelity, with the graph correctly holding its own bug-fix story (the decision, the dead-end it replaced, and the constraints, all wired). Search on that graph: on-topic queries returned 15/15 relevant hits; off-topic queries come back explicitly labeled weak — the graph likely has nothing on this rather than posing as answers.

Self-maintenance

The graph keeps itself clean server-side (no agent effort):

• Detect ($0): duplicate blocks, fabricated source quotes, schema drift.
• Resolve: a server-side reviewer auto-merges the clear cases; the genuinely ambiguous ones it can't decide are surfaced to the agent (it pulls them in workspace_stats) to confirm in plain language.

Infrastructure

Backup: automatic (after reflect cycles, keeps recent snapshots in ~/.nodedex/data/backups/); manual via POST /api/admin/backup; export via GET /api/export.

Documentation

See docs/README.md for the full index.

License

NodeDex is licensed under the GNU AGPL-3.0 (see LICENSE). You can use, modify, and self-host it freely. If you modify NodeDex and offer it to others over a network, the AGPL requires you to make your modified source available under the same license.

Commercial license: if you need to use NodeDex without the AGPL's copyleft obligations — e.g. embedding it in a closed-source product, or offering it as a hosted service without releasing your changes — a commercial license is available. Contact [email protected].

Contributions are welcome under a lightweight CLA — see CONTRIBUTING.md.