mindgap

Give an agent a goal and a place to remember. It loops — reading what's known, researching, writing back evidence-linked findings — and a knowledge graph grows itself across sessions. mindgap is that memory: a local, org-roam-style graph for research and project knowledge (concepts, definitions, software, repos, Confluence pages, arXiv papers, people, teams) that autonomous loop sessions read before they work and write after — markdown nodes densified with [[wiki-links]], every node carrying its source URLs, rendered live in 2D/3D. You can hand-curate via CLI + web UI too; the agents just never stop adding.

The idea

"LLMs are exceptionally good at looping until they meet specific goals. Don't tell it what to do — give it success criteria and watch it go." — Andrej Karpathy

An agent looping toward a goal needs somewhere to look before it starts and somewhere to put what it finds. mindgap is that somewhere. Each session reads the relevant subgraph for context, does the work — sweep arXiv, map a repo, mine the connections for buildable ideas — then ingests new nodes and edges with provenance. Nothing evaporates when the context window closes: the next run builds on the last, and knowledge compounds instead of being re-derived. You supply the goal and the success criteria; the graph is the durable, queryable memory the loop reads and writes.

Related MCP server: Kremis

Quickstart

pipx install git+https://github.com/grburgess/mindgap.git
mindgap init      # create ~/.mindgap/mindgap.db + a small demo graph
mindgap serve     # open the web UI at http://localhost:8765

The seeded graph is just a demo to show the shape. To start your own, clear it — the database recreates itself empty on next use:

rm ~/.mindgap/mindgap.db
mindgap add --title "My first concept" --tags research   # then grow it via the CLI, the web UI, or MCP

(Don't re-run mindgap init after clearing — that re-seeds the demo. Or point MINDGAP_DB at a new path to keep the demo and start a separate graph.)

Stdlib-only Python 3.10 — no pip installs. Data lives in a single SQLite file.

Install

Pick one. All paths put mindgap (and mindgap-mcp) on your PATH and store data in ~/.mindgap (override with MINDGAP_HOME, or MINDGAP_DB for just the DB file).

pipx (recommended): pipx install git+https://github.com/grburgess/mindgap.git mindgap init # creates ~/.mindgap/mindgap.db and seeds it mindgap serve # web UI at http://localhost:8765

pip: pip install --user git+https://github.com/grburgess/mindgap.git mindgap init && mindgap serve # ensure ~/.local/bin is on PATH

From a clone (no install / development): git clone https://github.com/grburgess/mindgap.git && cd mindgap ./install.sh # self-locating: PATH + ~/.mindgap + seed mindgap serve

Claude Code plugin (skills + MCP)

/plugin marketplace add grburgess/mindgap
/plugin install mindgap

Registers the mindgap MCP server and the paper-to-mindmap, arxiv-explainer, papers-library, and loop-system skills. Register the MCP at user scope so every Claude Code session, in any directory, can reach the graph:

claude mcp add -s user mindgap mindgap-mcp   # global; needs mindgap-mcp on PATH

-s user is what makes it global (the default scope is local/current-dir only). The launcher self-locates and the DB lives in ~/.mindgap, so it runs from anywhere. A source checkout also ships a project-scoped .mcp.json → ./bin/mindgap-mcp, active only inside this repo.

Where data lives

~/.mindgap/ — mindgap.db and snapshots/. MINDGAP_HOME relocates the whole dir; MINDGAP_DB points at a single DB file elsewhere.

Give an agent a goal

You don't drive this graph node-by-node — you point an agent at a goal and let it loop. Install the Claude Code plugin, then in any project just say:

"set up an arxiv-weekly loop watching <your topics> and run the first pass"
"continue the <name> loop"
"ideate buildable implementations from my <name> graph, and refute the ones that aren't feasible"
"build a graph of the authors doing <your topics> work, with their github pages"

Every run follows the protocol in AGENTS.md — read the existing subgraph for context, research, then ingest new nodes/edges with provenance and [[wiki-links]], and export a snapshot. The agent reaches the graph through the MCP server (validated writes that can't silently desync) or the CLI below. Beyond loops, every Claude Code session can deposit what it learned automatically. Then watch it compound — and re-run the loop tomorrow to grow it further. The pieces:

• Agent loops — give a goal (a topic, a library, a question); the loop runs to it and writes findings back, self-tuning each pass.

• MCP server — the read/write interface agents use, with guardrails (no partial commits, no dangling edges, provenance required).

• Self-learning capture — a SessionEnd hook that distills on-domain learnings from any session into the graph, unattended.

CLI cheatsheet

mindgap add --title T [--id ID] [--type TYPE] [--body MD | --body-file F]
            [--tags a,b] [--url KIND=URL ...] [--by AGENT]
mindgap link SRC DST [--rel REL] [--weight W] [--by AGENT]
mindgap ingest FILE|-                       # bulk JSON ('-' = stdin)
mindgap find QUERY [--type T] [--tag T] [--json]
mindgap show ID [--json]                    # node + neighbors + urls
mindgap context QUERY [--depth 1]           # markdown digest (for agents)
mindgap rm ID
mindgap unlink SRC DST [--rel REL]
mindgap export [--out FILE]                 # JSON snapshot -> ~/.mindgap/snapshots/
mindgap stats
mindgap lint [--json]                       # graph health: orphans/stubs/dups/stale
mindgap serve [--port 8765] [--no-open]

Web UI

mindgap serve opens a single-page graph viewer (dark editorial theme), in 2D or 3D:

• Force layout that spreads out instead of clumping — charge repulsion, a per-node collision force (2D), and tuned link distance keep nodes from overlapping, and the view auto-fits whenever the layout settles. Nodes are colored by type and sized by degree; hover a link to see its rel.

• Settings drawer (gear, top of the header) for live tuning, persisted to localStorage: a dark-theme picker (Editorial, Midnight, Graphite, Aubergine, Carbon), repulsion, link distance and strength, collision, link opacity, arrows, label mode, and the cluster controls below. "Reset to defaults" restores everything.

• Cluster feedback. Idea-communities are detected with multi-level Louvain — client-side, deterministic, no extra endpoint. Switch node coloring from type to community to surface them; in 2D each community gets a translucent hull and a centroid label, and in 3D each community gets a glowing nebula orb, with its topic name shown when you hover the orb. A legend (bottom-left) lists every community with its size — click one to isolate it and dim the rest. A Topic repulsion toggle adds a cohesion force that pulls each community toward its own centroid, so topics settle into spatially separate regions (works in 2D and 3D).

• Labels in four modes (off / hover / hubs / always). In hubs mode the most-connected nodes stay labelled and the rest fade in as you zoom into a region (2D); 3D shows a name on hover. Topic (community) names appear when hovering a 3D topic orb.

• Star field (3D). A subtle, twinkling star field sits behind the 3D scene (toggle in settings). It lives in world space, so orbiting and zooming give real depth parallax — the graph becomes a galaxy you drift through.

• Search box and type/tag filter chips narrow the graph live; a stats line sits in the header.

• Click a node → sidebar with its markdown body (wiki-links are clickable), tags, outbound URLs (open Confluence/GitHub/arXiv in a new tab), and neighbors. Edit body/tags/urls, link to another node via a search picker, or delete — all inline.

• Focus mode (double-click a node) shows its local 1-hop graph; clicking nodes, wiki-links, or neighbors while focused spreads the view ring by ring (org-roam style), and Esc or "unfocus" resets. Selecting any node flies the camera to it and highlights its neighborhood.

• Timeline (clock toggle in the header) opens a bottom strip: a per-day histogram of when nodes were added, a draggable playhead + ▶ play that grows the graph over time, and a color by recency / provenance toggle. Backlinks — the sidebar shows linked mentions plus unlinked mentions (nodes whose text names this one) with a one-click link. Quick switcher (Cmd/Ctrl-O) fuzzy-jumps to any node. Orphans chip filters to disconnected nodes.

The UI is vanilla JS with no build step, drawing force-graph/3d-force-graph, d3, marked, and dompurify from CDNs. Community detection and hull geometry live in web/cluster.js; the 3D topic glow (nebula orbs + orb-hover labels) and the twinkling parallax star field live in web/glow3d.js and web/starfield.js, which build sprites/points in the live Three.js scene using a version-pinned three ESM import (esm.sh/three@0.179) exposed as a global.

A tour of the UI

Topic clusters (2D). Color nodes by community, then flip on Topic repulsion — a cohesion force pulls each topic into its own region.

3D mode. The same graph in three dimensions — drag to orbit, scroll to zoom. A twinkling, world-space star field sits behind the scene, so the graph reads as a galaxy you drift through.

Dark themes. Five built-in dark themes — Editorial, Midnight, Graphite, Aubergine, Carbon — switched live.

Timeline. Scrub the playhead or hit ▶ to watch the graph grow over time — at day / week / month resolution, with a before / after toggle to show only what existed then or only what's new since.

Quick switcher. Cmd/Ctrl-O to fuzzy-jump to any node by title.

MCP server

For agents, mindgap/mcp.py exposes the graph as an MCP server over stdio — stdlib-only (newline-delimited JSON-RPC 2.0, no pip deps). For all sessions everywhere, register it globally at user scope: claude mcp add -s user mindgap mindgap-mcp (needs mindgap-mcp on PATH; the launcher self-locates and the DB lives in ~/.mindgap, so it works from any directory). A source checkout also ships a project-scoped .mcp.json → ./bin/mindgap-mcp, active only inside the repo.

Ten tools wrap the same db layer as the CLI: mindgap_ingest (batch write), mindgap_add_node, mindgap_link, mindgap_unlink, mindgap_get_node, mindgap_find, mindgap_context, mindgap_stats, mindgap_export, mindgap_remove_node. Unlike the raw CLI, the write tools validate at the call boundary — mindgap_ingest rejects the whole payload (no partial commit) if any edge endpoint isn't in the DB or the payload, mindgap_link refuses to auto-stub a missing endpoint, created_by is required, and writes return the persisted rows so a caller can't claim a write that didn't land.

Agent loops

The graph is designed to be fed by recurring autonomous sessions that scan Confluence, GitHub, and arXiv. The protocol — read context first, ingest JSON with provenance (created_by, source URLs), wiki-link into the existing graph, export at session end — is defined in AGENTS.md. Sessions can drive the graph via the CLI or the MCP tools above (the MCP's validation makes it the safer path for unattended writes).

Self-learning capture

Disabled by default. mindgap ships the capture engine off, with an empty domain. Nothing fires until you opt in.

Optionally, mindgap can learn from every Claude Code session: a SessionEnd hook runs a cheap deterministic pre-gate (no LLM) and, only when a session looks on-domain, fire-and-forgets a detached headless subagent that distills durable learnings and ingests them — following the knowledge-capture skill and AGENTS.md. Captured nodes carry created_by="capture:<repo>", confidence=0.6, and a urls entry pointing at the transcript, so they sit below hand-curated nodes and are trivially reversible.

To enable it:

1. mindgap init once — copies the packaged preset to ~/.mindgap/capture.json.

2. Edit ~/.mindgap/capture.json: set "enabled": true and fill in domain (a description and keywords that define what counts as on-topic). Tune denylist_dirs/allowlist_dirs, min_transcript_bytes, and the capture/lint blocks as needed. (MINDGAP_CAPTURE_ENABLED env-overrides the flag.)

3. Register the hook globally in ~/.claude/settings.json under SessionEnd, pointing at mindgap-capture-hook (on PATH after install, or ./bin/mindgap-capture-hook from a source checkout).

The pre-gate skips off-domain dirs, denylisted dirs, capture's own self-spawned sessions (MINDGAP_CAPTURE=1), too-small transcripts, and transcripts with no domain keywords — so the LLM subagent only ever runs on genuinely on-topic sessions. A best-effort lock (~/.mindgap/capture.lock) single-flights it. The hook never blocks session exit.

mindgap lint is the companion: a deterministic health report (orphans, dangling stubs, near-duplicate candidates, stale capture nodes) that never rewrites the graph.

Knowledge loops (arXiv search → graph)

The bundle ships self-improving loop templates that sweep arXiv for a topic and ingest findings into your graph with evidence-backed links — driven by the loop-system skill.

List what's available and scaffold one: mindgap loop list mindgap loop new arxiv-weekly --name my-watch --topics "your research area"

Then just tell Claude (in the project where you scaffolded it): "continue the my-watch loop"

Bundled templates:

• arxiv-weekly — recurring weekly 7-day arXiv sweep; tags every find, self-tunes its query strategy each pass. Schedule it unattended via the generated CRON.md (launchd/cron).

• paper-discovery — one-shot batch discovery of papers for a topic.

• paper-links — densify the graph by finding missing links between existing papers.

• implementation-ideation — mine the graph's growing connections for buildable ideas, then adversarially refute the infeasible ones; only vetted ideas (each with an MVP sketch) are ingested.

• author-graph — build a person-node graph of the researchers behind the work, with their resolved GitHub / homepage / Scholar links and co-author connections.

Share a loop you've built (strips your accumulated state): mindgap loop export my-watch # -> ./my-watch-template/ mindgap loop import ./my-watch-template --name their-watch --topics "..."

Prompts you can hand to Claude directly (once the plugin is installed):

• "set up an arxiv-weekly loop watching and run the first pass"

• "continue the loop"

• "export the loop as a template I can share"

• "ideate implementations from the growing connections in my graph, and refute the ones that aren't feasible" (implementation-ideation)

• "build a graph of the authors doing work, with their github pages" (author-graph)

Paper explainers

The bundled arxiv-explainer skill turns a paper into a richly animated, narrated HTML explainer — figures extracted from the PDF, a self-contained dark theme — and ingests it into your graph. Just tell Claude explain <arXiv link> (or point it at a local PDF).

Import a Papers library

Mine your Papers (ReadCube) reference library into the graph: export it to BibTeX or RIS (Papers → Settings → Export) and tell Claude:

"import my Papers library from <path-to-export.bib>"

The bundled papers-library skill parses the export (stdlib, no deps), ingests each paper as a node (deduped against the graph, evidence-linked), discovers related papers not yet in your library, and seeds ideas — handing off to the paper-links / implementation-ideation loops for depth.

Schema overview

Two tables:

• nodes(id, title, type, body, tags, urls, confidence, created_by, created_at, updated_at) — id is a kebab-case slug; tags/urls are JSON arrays; types: concept|definition|software|repo|page|paper|person|team|stub.

• edges(src, dst, rel, weight, created_by, created_at) — rels: relates_to|defines|implements|depends_on|cites|part_of|mentions.

[[wiki-links]] in a body sync to mentions edges automatically, creating stub nodes for missing targets. Upserts merge: scalar fields replace, tags/urls union.

Export & snapshots

The DB is gitignored; history is kept as JSON snapshots:

mindgap export                # ~/.mindgap/snapshots/<utc>.json
mindgap export --out my.json

Commit snapshots for a durable, diffable record; re-ingest one with mindgap ingest FILE to restore.

Development

python3 -m mindgap ...                 # run CLI from repo without install
python3 -m unittest discover tests

License

MIT — see LICENSE.