Hypermnesia Mcp Viz

[Hypermnesia MCP Viz, the read-only visualization layer for Cortex: six live reading angles over your memory, sessions, and code, rendered but never remembered]

[Hypermnesia MCP Viz Trace galaxy on the paper surface: every Claude Code session is a tight phyllotaxis disk of its own prompt → action → file → memory chain, clustered around its domain's olive hub; the selected chain's disks render solid gold in place while the rest of the galaxy stays legible, and the mono status bar streams the exact counts (CHAIN · 108 STEPS · 6 537/6 537 nodes · 8 696/8 696 edges), never rounded.]

[Cross-platform MCP for Codex, Gemini CLI, and Claude Code] [MIT License] [Python 3.10+] [Version 3.2.0] [OpenSSF Best Practices]

Hypermnesia MCP Viz

Hypermnesia MCP Viz is published as hypermnesia-mcp-viz everywhere: Python, MCP Registry, Claude Code, Codex, Gemini, and the console command.

The cross-platform visualization layer for Cortex. Codex, Gemini CLI, Claude Code, and any stdio MCP host can launch the same read-only server and browser UI. It turns Cortex's memory store, a host-neutral live activity stream, Claude Code's historical session archive, and your codebase graph into six live reading angles: a project galaxy, a 3D anatomical brain, an execution trace, a consolidation kanban, a curated knowledge browser, and a wiki. It never writes a memory: it renders, it never remembers. (It does keep its own derived graph/layout caches in the Cortex database, listed under Boundary.)

Launch with the open_visualization tool (or /cortex-visualize). One launcher opens six reading angles; the default landing view is Trace.

The whole UI ships on the AI Architect design system: a paper-first reading surface with a persistent ink (night) toggle, greyscale chrome, and a data palette resolved from design tokens at runtime, flipping the surface re-inks every view in place without disturbing a settled layout, and every count on screen is exact and streamed, never rounded.

Getting Started

Claude Code users can install Hypermnesia MCP Viz from the same cortex-plugins marketplace as Cortex:

claude plugin marketplace add cdeust/Cortexclaude plugin install hypermnesia-mcp-viz@cortex-plugins

This install name becomes available in the Cortex marketplace only when this repository and the matching marketplace change in cdeust/Cortex#351 are released together. Maintainers must not publish either half of the rename on its own.

Migrating from cortex-viz

Existing Claude Code installs are not aliases for the new plugin identity. Remove the old install, refresh the marketplace after the coordinated release, and install the canonical plugin:

claude plugin uninstall cortex-viz@cortex-pluginsclaude plugin marketplace update cortex-pluginsclaude plugin install hypermnesia-mcp-viz@cortex-plugins

Claude composes plugin tool names from the plugin manifest name, the mcpServers key, and the tool name. Update both old full names wherever they appear in permission allowlists, hooks, skills, or agents:

• mcp__plugin_cortex-viz_cortex-viz__open_visualization becomes mcp__plugin_hypermnesia-mcp-viz_hypermnesia-mcp-viz__open_visualization.
• mcp__plugin_cortex-viz_cortex-viz__get_methodology_graph becomes mcp__plugin_hypermnesia-mcp-viz_hypermnesia-mcp-viz__get_methodology_graph.

The removed cortex-viz console command is not retained as an alias, so Cursor, Windsurf, VS Code, and other direct-process configurations must invoke hypermnesia-mcp-viz instead.

Hypermnesia MCP Viz is a companion to Cortex that never writes a memory. Install Cortex first (claude plugin install hypermnesia-mcp@cortex-plugins): the visualizer reads its shared PostgreSQL store, and writes only its own derived caches there (Boundary). Point both at the same database: the database_url plugin setting defaults to postgresql://127.0.0.1:5432/cortex; set it to the same value you gave Cortex.

Restart your Claude Code session, then launch the visualizer:

/cortex-visualize

Codex, Gemini CLI, and other MCP hosts launch the same server as a plain stdio process; see Other MCP hosts for the exact configuration.

One launcher opens all six reading angles (Graph · Brain · Trace · Knowledge · Wiki · Board) in the browser, served live from the Cortex store, your session JSONL, the code graph, and git.

Works without Cortex

No Cortex, no PostgreSQL, no setup: Hypermnesia MCP Viz is still useful on its own. The Trace view (the default landing view) reads only ~/.claude/projects/*.jsonl and your local git: every Claude Code session becomes a navigable domain → session → prompt → action → file chain, with per-file diffs and commit history. If you use Claude Code, the data is already on your disk.

• Just install the plugin and run /cortex-visualize. When Cortex's database isn't reachable, the server logs one line and starts in no-DB mode automatically: Trace is fully live; the five DB-backed views (Graph, Brain, Knowledge, Wiki, Board) appear greyed out with an install pointer instead of erroring.
• To skip the database probe entirely, set CORTEX_VIZ_NO_DB=1 (or pass --no-db when running the standalone server directly).

Installing Cortex later lights up the other five views against the same UI (no reconfiguration).

git clone https://github.com/cdeust/cortex-viz.git && cd cortex-vizpip install -e .DATABASE_URL=postgresql://127.0.0.1:5432/cortex python3 -m cortex_viz# Without Cortex/Postgres (Trace view only):CORTEX_VIZ_NO_DB=1 python3 -m cortex_viz

The views

Graph: the Claude workflow map

Each project becomes a cloud of nodes around one gold domain hub. Inside every cloud, nodes sit in six concentric levels by the Claude surface (or the code itself) that produced them:

Why L6 matters. L5 and below tell you what Claude did; L6 tells you what the code is. Three things become visible for free: shared code (any symbol referenced by two projects drifts into the inter-project gap), impact (clicking a symbol surfaces every caller, importer, and member: "what breaks if I change this?" is a graph neighbourhood, not a grep), and the shape of the codebase itself (a dense petal around a file means a fat internal API; a thin one means a leaf module). A grouped filter (L1–L6 / by kind / by AST edge kind / Cross-domain) isolates any slice.

Brain: the galaxy inside a real cortex

The same graph, on a second surface: every node placed inside an anatomical cortical mesh by the neuroscience of memory systems rather than by force-direction. Episodic memories sit in the medial temporal lobe and migrate outward to neocortex along a hot→consolidated depth gradient (the complementary-learning-systems consolidation model); semantic entities in temporal neocortex; code symbols in association cortex; procedural skills in the striatum and cerebellum; domains at the connectome's rich-club hubs. Region centres are registered from real MNI152 atlas coordinates (affine fit, not vertex-exact; the mesh is a single unlabeled surface). Every synapse routes along a major white-matter tract (fornix, uncinate, SLF, corpus callosum). Node colour is the same semantic palette as the galaxy (memories by consolidation stage, entities/symbols by type), and a live Memory science panel mirrors the store's system vitals (consolidation pipeline, skills, source-monitoring, extinction, sleep phases, and every mechanism Cortex exposes).

Because the full graph (278k+ nodes, 5.5M edges) is far larger than a browser can take in one payload, the brain streams it in progressively through a bounded-queue, frame-budgeted NDJSON loader: the cloud fills in as you watch. Clicking any node opens the same rich detail card as the galaxy (content, tags, live heat, relations, git diff, impact). Open it from the Brain button in the view bar, directly at /brain, or programmatically via open_visualization(view="brain").

[3D Brain view: 125,744 nodes streamed into a translucent anatomical cortical mesh; episodic memories in the medial temporal lobe fading outward to neocortex along a hot→consolidated depth gradient, semantic entities and code symbols in association cortex, procedural skills toward the striatum and cerebellum, domain hubs at the connectome's rich-club centres; the left panel reads node / synapse / domain / memory / entity counts plus memory-science vitals (mean heat, growing, strong, conflicts, forward-model error) and the right panel legends the memory-system regions and exact per-category node counts]

[Board view: five columns for labile, early LTP, late LTP, consolidated, and reconsolidating memories on the paper surface; each column header states the exact live count (117 · 2841 · 41 · 1 · 0) with its reference stage physics (decay ×, vulnerability %, plasticity %, from the DD-02 spec) and the factual advance rule; below, domain and feeling facet chips with exact counts filter the cards, and the empty reconsolidating column shows an honest 'No memories' instead of hiding]

Board: consolidation as a kanban

Five columns by consolidation stage (labile · early_ltp · late_ltp · consolidated · reconsolidating). Each header states the stage's live count (server-side truth, not the loaded page) above its reference stage physics (decay ×, vulnerability %, plasticity %) and the stage's advancement rule (e.g. replays ≥ 1 or imp > 0.3 at labile, replays ≥ 3 at early_ltp). Those physics rows and advance rules are constants from the DD-02 stage spec, not per-bucket medians recomputed live. Cards carry live heat, importance, surprise, valence, arousal, and the exact tool that created the memory.

Detail panel: every measurement explained. Clicking any node opens a panel with the raw value and a one-line plain-language explanation. Consolidation stage, activity (heat), importance, surprise, emotional tone and intensity, confidence, plasticity, stability: each a labeled bar with a sentence like "How unexpected this memory was when it arrived. Surprises stick better than routine events."

[Trace view, session drill-down: clicking a session opens its full conversation replay: the transcript modal streams every assistant message and tool call (WRITE · READ) with timestamps, down to the closing checkpoint the agent wrote; the right panel holds the session's discussion card, and the mono status bar keeps the exact chain and node/edge counts]

Trace · Knowledge · Wiki

• Trace (default): the live execution-trace drill: collapsed domain hubs → sessions → the ordered prompt → action → file chain of what actually happened → a file's AST symbols, impact neighbourhood, and git history. Discussions and Cortex remember/recall ops are woven into the chain. Served live from session JSONL, the code graph, and git on every request (no snapshots, always current).
• Knowledge: curated memory cards with the feeling (word + signed valence/arousal, never colour alone), the MEANING line and verbatim excerpt, stage/domain/HOT badges, and four measured meters in fixed order (heat · importance · valence · arousal; a zero shows an empty track, never hides); filter by domain, stage, or feeling with exact facet counts.
• Wiki: the per-project knowledge base as a browsable Project → Kind → Pages tree with a dossier-style page reader: serif prose with numbered section heads and mono identifier chips, boxed status and kind badges, dated provenance, and an Edit · PDF · TEX · DOCX · HTML export strip. A CodeMirror split-pane editor with live preview sits behind Edit. (The wiki content is authored autonomously by Cortex; Hypermnesia MCP Viz is its reading + editing surface.) Pages Cortex publishes into a repository's own wiki/ (its project_root mode, recognised by wiki/manifest.json) are listed under that project with a @<project>/ path prefix; they are read-only here, since Cortex publishes each with its manifest and docs/adr mirror.

[Knowledge view, curated memory cards on the paper surface: each card carries its tool source, feeling word with signed valence/arousal deltas, MEANING line with verbatim path, stage badge (LABILE · EARLY-LTP), domain chip, HOT flag, four measured meters (heat, importance, valence, arousal), and capture provenance; the filter bar above states exact counts per domain, stage, and feeling]

[Wiki page reader: REFERENCE / AGENTIC-AI breadcrumb over the serif title 'agentic-ai: MCP integration' with boxed DRAFT and REFERENCE badges, CREATED · UPDATED provenance in mono micro-caps, an Edit · PDF · TEX · DOCX · HTML export strip, the project tree rail (Tree/Graph segmented control, per-kind page counts, accent-ticked active page), and a serif prose body with mono identifier chips and numbered section heads]

Install

Hypermnesia MCP Viz is a cross-platform MCP server with a Claude Code plugin. Point it at the same database as your Cortex install: it reads Cortex's memories from that store and never writes them.

As a plugin: ships the MCP server, the /cortex-visualize skill, and the live session-activity hooks. The bundled scripts/launcher.py bootstraps its own dependencies on first launch (no manual pip needed). Configure the DB via the plugin's database_url user-config (defaults to postgresql://127.0.0.1:5432/cortex).

As a raw MCP (v3.1.0 and later):

pip install "hypermnesia-mcp-viz[data,viz-tile]"   # optional PG and large-graph extrashypermnesia-mcp-viz                 # or: python -m cortex_viz (stdio MCP transport)

The published 2.8.0 hypermnesia-mcp-viz distribution owned the legacy cortex-viz console shim; there was no separate Python distribution by that name. Upgrade with the same interpreter/environment that installed it, and the installer removes the old distribution-owned shim while installing 3.1.0 (the distribution-identity rename that removes the shim was cut into the tree as 3.0.0, but that version was never tagged or published — see the CHANGELOG's [3.0.0] entry — so 3.1.0 is the first published release that carries it):

python3 -m pip install --upgrade "hypermnesia-mcp-viz>=3.1.0"

If cortex-viz still resolves, first locate every shim and ask Python which installed distribution owns that console entry point:

type -a cortex-vizpython3 - <<'PY'from importlib import metadata
owners = []for distribution in metadata.distributions():    if any(        entry.group == "console_scripts" and entry.name == "cortex-viz"        for entry in distribution.entry_points    ):        owners.append((distribution.metadata["Name"], distribution.version))print(owners or "no Python distribution owns cortex-viz in this interpreter")PY

Uninstall the reported distribution with that same interpreter, or remove a manually installed shell/file shim through the tool that created it when no distribution owns it. Do not guess a distribution name from the executable.

Set DATABASE_URL to the shared Cortex database. open_visualization launches the galaxy UI in the browser, bound to 127.0.0.1.

Other MCP hosts

Any MCP host can launch the server: it is a plain stdio process. Install it with pip install hypermnesia-mcp-viz, then register hypermnesia-mcp-viz as both the server name and command in your host's MCP config (Gemini CLI ~/.gemini/settings.json, Cursor .cursor/mcp.json, Windsurf ~/.codeium/windsurf/mcp_config.json, VS Code .vscode/mcp.json under "servers", or Codex CLI: codex mcp add hypermnesia-mcp-viz -- hypermnesia-mcp-viz). The equivalent Python module form is python3 -m cortex_viz. The open_visualization tool opens the UI in the browser. The historical Trace archive still reads Claude Code JSONLs under ~/.claude/projects/, but live activity is host-neutral: a Codex, Gemini, or generic MCP adapter can POST the versioned schema below to the local /api/activity endpoint. The galaxy/brain/knowledge/wiki/board views still need a Cortex store to read.

Host-neutral live activity

POST /api/activity accepts the JSON Schema contract in docs/host-event-v1.schema.json alongside the existing Claude hook payload. Producers identify their host explicitly; tool names are provenance and are not reinterpreted as Claude tools.

{  "schema_version": "1",  "host": "codex",  "session_id": "01J...",  "timestamp": "2026-08-02T12:34:56Z",  "event": "tool_call",  "tool": "read_file",  "input_summary": "Read the authentication module",  "artifact": "src/auth.ts",  "result": "success",  "cwd": "/workspace/project"}

The endpoint normalizes the event into the same session/action/target graph as Claude activity. Producers can report mcp_call, api_call, db_read and db_write explicitly in addition to prompts, tools, files, terminal commands, skills, subagents and web access. API/database targets come from the producer's artifact; input/result summaries are displayed as observed and are never inferred from a tool name. Host-specific hook or telemetry adapters can be added independently without changing the graph or database schema.

Boundary

Hypermnesia MCP Viz consumes Cortex's artifacts on disk + PostgreSQL, never Cortex's live Python objects:

No import mcp_server.* is permitted anywhere in cortex_viz/: that invariant is the extraction's correctness check.

MCP tools

open_visualization (launch the browser UI: pass view="brain" for the 3D anatomical brain, view="galaxy" or omit for the 2D graph) and get_methodology_graph (graph data). The six views are served over HTTP by the server open_visualization launches; a live session-activity stream (every tool call, MCP call, file access, skill, and command) feeds the graph in real time via the activity-capture hooks.

Project

Achievements

• OpenSSF Best Practices Silver: project 13846 earned Silver on 2026-08-03. Answers for every passing and Silver criterion are committed in .bestpractices.json and audited against this repository rather than copied.
• OpenSSF Scorecard: runs weekly (scorecard.yml). Current score 7.4 (2026-08-03), including the Silver badge and zero known vulnerabilities.
• Build provenance: all four v2.8.0 artifacts — wheel, source archive, CycloneDX SBOM, and UI manifest — have independently verified Sigstore-backed attestations. Verify any downloaded artifact with gh attestation verify <file> --repo cdeust/cortex-viz.

Status

The visualization stack was extracted from Cortex (which is now a focused memory engine) so the graphics ship and scale on their own. Standalone MCP boots over stdio; all six views are bridged to live data; the galaxy builds end-to-end at 75k+ nodes; the 3D brain streams the full graph into a cortical mesh; the whole UI sits on the AI Architect design system (paper/ink surfaces, token-resolved palette); the suite passes.