odysseus/docs/memory-graph-analysis.md

Memory Graph View — Repository Analysis

Status: research only. No application code, dependencies, or database state was modified to produce this document.

Scope: this document inventories the parts of the Odysseus codebase relevant to building an interactive, Obsidian-like Memory Graph View, and identifies the safest points to extend the system. A companion document, docs/memory-graph-design.md, proposes the actual design based on these findings.

1. Frontend architecture

• No bundler, no build step. static/js/package.json sets { "type": "module" } so the browser loads native ES modules directly. There is no webpack/vite/esbuild config anywhere in the repo. The root package.json has no scripts block; its only devDependency is @antithesishq/bombadil, a browser-fuzzing spec library unrelated to bundling.
• Entry point: static/index.html loads static/app.js via <script type="module">. app.js statically imports ~30 feature modules, each a plain ES module with a default export object (e.g. import memoryModule from './js/memory.js?v=20260722memoryloading1'). Cache-busting is done manually via query-string suffixes on the import path.
• Cross-module reachability: some modules are also hung off window (window.themeModule, window.sessionModule, window.uiModule, window.adminModule, window.cookbookModule) so unrelated modules can call into them without a formal event bus.
• static/js/ has ~100 flat files plus subfolders for cohesive subsystems: editor/ (canvas-based image editor), compare/, research/, calendar/, emailLibrary/, markdown/, model/, color/, util/. static/js/MODULE_SUMMARY.md is the authoritative, actively maintained architecture index and should be updated alongside any new module.
• Backend communication is plain fetch() returning JSON, plus one Server-Sent Events (SSE) stream for chat (chat.js posts to /api/chat_stream, reads via res.body.getReader(), parses data: {...} lines by type). No WebSockets exist anywhere in the codebase.
• app.js patches window.fetch globally so any 401 response redirects to /login — any new module's fetches automatically inherit this behavior.

2. Routing

• There is no client-side SPA router (no history-based route table, no hash router library). Odysseus is a single persistent DOM (index.html) where "navigation" means opening/closing modals and full-screen panels, not swapping views by URL.
• A lightweight deep-link opener exists in app.js: a _routeOpen map keyed by window.location.pathname (entries for /notes, /calendar, /cookbook, /email, /memory, /gallery, /tasks, /library). The server presumably serves index.html for these paths (catch-all), and this map simulates the click that would normally open the corresponding modal. A new /memory-graph entry point would follow this exact pattern, or the graph could live as a new tab inside the existing memory modal instead of a new top-level route.
• Everything else (Memory, Calendar, Gallery, Documents, Tasks, Compare, Cookbook, Settings) is a .modal element in index.html toggled by modalManager.js. There is no hash-based view dispatch beyond an ad hoc entity-hash regex in init.js used only for composer-restore behavior.

3. UI component library / design system

• No component framework (no React/Vue/etc.), no CSS framework (no Tailwind/Bootstrap). static/style.css is a single hand-written stylesheet (tens of thousands of lines) using CSS custom properties for theming (:root { --bg; --fg; --red; ... }), a :root.light override, density variants (.density-compact, .density-spacious), and a UI-scale zoom mechanism.
• The "component" pattern is plain JS factory functions building DOM via document.createElement and manual event wiring — no virtual DOM, no templating engine. static/js/memory.js builds each memory-list row this way.
• Modals share a common infrastructure:
  • static/js/modalManager.js — central open/minimize/close/dock manager. Public API: Modals.register(id, { railBtnId, restoreFn, closeFn }), Modals.toggle(id). Owns a draggable "minimized dock" tray with FLIP animations and magnetic close-on-drag-to-trash behavior. Each modal type has a label+icon entry in _LABELS, including an existing 'memory-modal': { label: 'Brain', icon: ... } entry. A new Memory Graph View, if its own modal/window, should register here to get consistent minimize/dock/restore/z-order behavior for free.
  • static/js/modalSnap.js — edge-docking/snap-to-zone logic, imported by modalManager.js.
  • static/js/tileManager.js — desktop window tiling/snap-to-edge, used by memory.js for snapModalToZone.
  • static/js/windowDrag.js — generic makeWindowDraggable(modal, opts) helper, used to make the memory modal's header draggable.
  • static/js/toolWindowZOrder.js — monotonically increasing z-index (nextToolWindowZ()) so the most-recently-focused tool window stacks on top.
• static/modal-control-variants.html, static/wave-variants.html, static/whirlpool-variants.html are standalone design-exploration/prototyping pages, not wired into the live app bundle.
• Third-party libraries that are needed client-side are vendored directly into static/lib/ (e.g. docx.umd.min.js, highlight.min.js, xlsx.full.min.js) rather than installed via npm — this is the established pattern for adding any graph-rendering library, since there is no bundler to run npm install through.

4. Existing memory UI

• static/js/memory.js (~1550 lines) implements the entire "Brain" modal. No graph, timeline, or relationship view exists today.
• Modal structure (index.html): #memory-modal with tabs — Browse, Skills, Add, Settings.
• Browse tab is a flat list view (#memory-list, .memory-item rows), not a card grid. Each row shows text, category badge, pinned badge, source (auto/manual), use-count, relative timestamp, and a kebab menu (Pin/Select/Edit/Delete).
• Supported interactions: free-text client-side search, category filter chips, sort dropdown (newest/oldest/A-Z/most-used) with pinned items floated to top, bulk multi-select with bulk-delete, inline double-click-to-edit, an AI-driven "Tidy" (dedupe/audit) action, import from file with LLM-extracted suggestion review, and JSON export.
• Categories are a fixed client-side set: MEMORY_CATEGORIES = ['fact','identity','preference','contact','project','goal','task'].
• The backend already exposes GET /api/memory/timeline, sorted by timestamp with resolved session names, but no frontend module calls it today — it is the closest existing "structured" memory endpoint and a plausible seed for a graph-view layout, but it is currently dead code from the UI's perspective.

5. Existing graph or visualization components

• None. An explicit search across static/, src/, routes/, and the whole repo for d3., cytoscape, vis-network, vis.js, sigma.js, force-graph, d3-force, three.js, and networkx returned zero matches.
• The only <canvas> usage in the codebase is the image editor (static/js/editor/*, static/js/galleryEditor.js) for pixel/layer compositing — unrelated to node-link rendering.
• Conclusion: a Memory Graph View is a greenfield UI addition. There is no library, canvas renderer, or "relations" concept in the data model to build on top of.

6. Backend architecture

• Framework: FastAPI (app.py), served by Uvicorn. Two entrypoints: app.py (standard dev/server) and launcher.py (Windows portable/frozen build with a tray icon, same Uvicorn server underneath).
• Route registration is manual, not auto-discovered: every routes/*.py module exposes a setup_*_routes(...) factory that builds and returns an APIRouter with dependencies passed as plain constructor args (not FastAPI Depends()). app.py explicitly imports and include_router()s over 40 of these factories.
• Startup uses the modern FastAPI lifespan context manager. src/app_initializer.py::initialize_managers() is a pure component factory invoked once at startup: it builds MemoryManager, SkillsManager, SessionManager, UploadHandler, PersonalDocsManager, APIKeyManager, PresetManager, MemoryVectorStore (Chroma-backed, degrades gracefully if unhealthy), wraps memory in a MemoryProviderRegistry, and builds ChatProcessor/ChatHandler/ModelDiscovery.
• Route handlers are async def, but database access uses classic synchronous SQLAlchemy (SessionLocal()), occasionally wrapped in asyncio.to_thread. The process is single-instance by convention (log rotation is explicitly not multi-process safe).
• Dependency injection is a manual "component bag" pattern, not Depends(): singletons are built once in app.py/app_initializer.py and closed over by each router factory. Request.state carries auth-derived values (current_user, api_token, etc.) set by middleware and read directly inside handlers.

7. API routes

• Auth is enforced by AuthMiddleware (inside app.py, conditional on AUTH_ENABLED=true) plus per-route calls into src/auth_helpers.py: get_current_user(request) (soft), require_user(request) (401 if unauthenticated), require_privilege(request, "can_manage_memory") (401/403 by privilege flag), effective_user(request) (resolves the real owner behind a Bearer API token).
• Ownership is enforced manually per route via helpers like _assert_session_owner/_verify_memory_owner in routes/memory/memory_routes.py, all raising 404 (not 403) on cross-owner access to avoid confirming another user's resource exists.
• Request/response validation is inconsistent by design across routes: some use Pydantic models from src/request_models.py (MemoryAddRequest, etc., with lenient field validators that clamp/default rather than hard-reject), others use raw Form(...) parameters. Pick per-route based on whether the caller posts JSON or a browser form.
• Error handling is two-layered: routes raise HTTPException directly and inline; a small set of domain exceptions (SessionNotFoundError, InvalidFileUploadError, LLMServiceError, WebSearchError) are registered globally in core/exceptions.py / app.py with fixed JSON shapes and status codes.
• Rate limiting (src/rate_limiter.py) is a simple in-memory sliding-window limiter, wired only into routes/auth_routes.py (login/setup) — there is no global rate-limit middleware.
• Pagination has no single shared convention. History uses offset/limit with a "default to most recent page" fallback; documents use Query(0, ge=0)/Query(20, ge=1, le=50); memory routes have no pagination at all — GET /api/memory and /timeline return the full owner-scoped list every time. A graph endpoint returning potentially hundreds of nodes/edges will need new pagination/limiting that doesn't exist in the memory API today.

8. Database schema

• Engine: SQLite by default (DATABASE_URL=sqlite:///{DATA_DIR}/app.db), synchronous SQLAlchemy declarative ORM. A custom EncryptedText TypeDecorator transparently Fernet-encrypts sensitive columns at the ORM layer.
• Migrations: no Alembic. Hand-rolled, idempotent _migrate_add_* functions run in a fixed sequence from init_db(), each checking PRAGMA table_info(table) before ALTER TABLE ... ADD COLUMN, safe to re-run every startup. init_db() also calls Base.metadata.create_all(bind=engine) for any brand-new tables.
• Key tables (all in core/database.py): sessions, chat_messages, memories (see discrepancy below), documents/document_versions, gallery_albums/gallery_images, notes, calendars/calendar_events/caldav_deleted_events, email_accounts, scheduled_tasks/task_runs, plus supporting tables (model_endpoints, api_tokens, mcp_servers, comparisons, signatures, webhooks, user_tools, crew_members, editor_drafts, integrations).
• Users/auth are not in the SQL database — core/auth.py::AuthManager reads/writes data/auth.json. Ownership on DB rows is a plain owner string column (username), not a foreign key to a users table; NULL conventionally means "legacy/shared, visible to everyone."
• Important discrepancy: a memories SQL table exists with proper schema and a session_id FK, but the runtime memory store actually used everywhere (routes, chat context injection, agent tool) is src/memory.py::MemoryManager, which persists to a flat JSON file data/memory.json, not this table. The SQL memories table appears to be unused/legacy. Any Memory Graph View must treat memory.json (+ its Chroma vector index) as the real source of truth, not the SQL table.

9. Memory system

• Canonical storage: src/memory.py::MemoryManager — a JSON file (data/memory.json) holding flat entries: id, text, category (default "fact"), source, owner, timestamp, optional session_id, pinned, uses. There is no relationship/edge field — no related_ids, links, or similar. Any graph view must derive edges rather than read stored ones.
• services/memory/memory.py and services/memory/memory_vector.py are thin backward-compatibility shims re-exporting the canonical src/memory.py / src/memory_vector.py implementations.
• Backend routes (routes/memory/memory_routes.py, prefix /api/memory; routes/memory_routes.py is a compat shim re-exporting the same router object): POST /add, GET "" (list), POST /search, GET /timeline, GET /by-session/{session_id}, POST /extract (LLM suggestion extraction from a chat session), POST /audit (dedupe), POST /import (file-based extraction), POST /{id}/pin, GET /{id}, PUT /{id}, DELETE /{id}.
• Memory-to-session relationship is informational provenance only (session_id records which chat session a memory was extracted from) and is not a structural graph edge.
• Two independent paths connect memory to the agent/chat pipeline:
  1. Automatic read path — src/chat_processor.py::ChatProcessor.build_context_preface(...) runs on every chat/agent turn when use_memory=True (default): loads the owner's memories, selects relevant pinned + hybrid-retrieved (BM25 + vector) memories up to a context limit, injects them as untrusted-context messages before the LLM ever sees the turn, and bumps each injected memory's uses counter.
  2. Explicit write path — the model can emit a manage_memory tool call (src/ai_interaction.py::do_manage_memory), supporting list|add|edit|delete|search, mutating MemoryManager + MemoryVectorStore directly and firing a memory_added event.
• src/memory_provider.py defines a MemoryProvider ABC / MemoryProviderRegistry intended to let alternate memory backends plug in behind the same interface; only NativeMemoryProvider is currently registered. This is the natural extension seam if a graph-capable provider is ever needed, though the design in this analysis targets the native provider directly since that's what all current UI and agent paths use.

10. ChromaDB integration / vector search

• src/chroma_client.py uses chromadb.HttpClient(host, port) — a real HTTP client against a separate chromadb container, not an embedded/local persistent client. Config: CHROMADB_HOST (default localhost), CHROMADB_PORT (default 8100); docker-compose sets these to chromadb/8000 for in-network container-to-container access. A TCP probe fails fast before constructing the client; client.heartbeat() verifies liveness before the singleton is cached.
• Collection names: odysseus_memories (src/memory_vector.py::MemoryVectorStore.COLLECTION_NAME) and odysseus_rag (src/rag_vector.py). Both are created lazily through build_embedding_lanes() → chroma_client.get_or_create_collection(...) (src/embedding_lanes.py), which supports multiple simultaneous embedding backends ("lanes": local FastEmbed ONNX, or a configured custom embedding endpoint), each with its own Chroma sub-collection, searched in parallel and de-duplicated.
• MemoryVectorStore exposes add(memory_id, text), remove(memory_id), search(query, k), and a healthy flag. This is the natural, already-available source of pairwise similarity scores for auto-generating graph edges between memories, since the memory schema itself has no explicit relation data.
• Memories and personal-doc RAG chunks live in separate Chroma collections and are not cross-linked or queried together today.

11. Document storage

• Two distinct subsystems:
  • "Living documents" (AI-editable canvas documents): routes/document_routes.py, backed entirely by the relational DB (documents/document_versions tables) — content lives in Text columns, not on disk. Document.owner is stamped independently of session_id so a document survives its parent session's deletion.
  • "Personal Docs" (RAG document library): src/personal_docs.py::PersonalDocsManager walks a directory on disk, extracts text (PDF via pypdf, Office via markitdown), chunks it, and indexes into the odysseus_rag Chroma collection. Files stay on disk; there is no separate SQL metadata table — the Chroma collection's metadata is the catalog.
• Not directly part of the Memory Graph View's data model, but a candidate future edge type ("memory ⟷ document it was extracted from") if the design is extended later.

12. Conversation storage

• Sessions and messages are relational DB rows (sessions, chat_messages), managed exclusively through core/session_manager.py::SessionManager.
• SessionManager keeps an in-memory cache but only loads the 100 most-recently-accessed, non-archived, non-empty sessions' metadata at boot (not messages) to bound memory; message history is hydrated on demand.
• routes/history/history_routes.py supports offset/limit paging directly against chat_messages, independent of the in-memory cache, and lazily "hydrates" a session's in-RAM history from the DB if it's found to be behind.
• sessions.owner (nullable username string) is the ownership authority that memory, documents, and other per-session resources ultimately trace back to via their own owner column or session_id FK.

13. Agent architecture

• src/agent_loop.py::stream_agent_loop(...) is the central async-generator agent loop: streams SSE events (delta, tool_start, tool_output, agent_step, metrics, [DONE]), assembles a dynamic system prompt, and handles plan-mode, tool policies, and per-model quirks.
• src/tool_execution.py dispatches model-emitted tool calls either to MCP servers (src/mcp_manager.py) or native Python implementations under src/agent_tools/. Sensitive filesystem tools are admin-only and path-confined via deny/allow lists.
• src/agent_runs.py lets an SSE stream survive a browser disconnect by draining the generator server-side into a replay buffer per session id (does not survive a server restart) — this is the closest existing pattern to a "live push channel," and is the template to follow if the graph view needs live updates (see §14).
• Memory interacts with the agent loop via the two paths described in §9 (automatic context injection, explicit manage_memory tool), not via any structural graph traversal today.

14. Real-time/event infrastructure

• src/event_bus.py::fire_event(event_name, owner) is a task-automation trigger bus, not a pub/sub-to-frontend mechanism. It matches ScheduledTask rows with trigger_type == "event" against the fired event name and runs the matching automation once a threshold count is hit. It has no subscriber API that a browser tab could listen to.
• Publishers of memory_added: routes/memory/memory_routes.py (on add), services/memory/memory_extractor.py, src/ai_interaction.py (memory captured during chat). The only built-in consumer is the "Memory Tidy" automation (fires a consolidate_memory task every 5 adds). There is currently no memory_updated or memory_deleted event fired anywhere.
• Genuine per-session pub/sub does exist, but scoped to a single request's own output stream: routes/chat_routes.py + src/agent_runs.py implement AgentRun.subscribers as a set of asyncio.Queue (one per connected client), fed into a StreamingResponse(media_type="text/event-stream"). Similar SSE streaming exists in routes/shell_routes.py, routes/model_routes.py, routes/research/research_routes.py.
• Implication: there is no ready-made mechanism today to push "a memory changed elsewhere" to an open Memory Graph View. The design doc proposes either polling or a new lightweight per-owner SSE channel modeled directly on the agent_runs.py queue-per-subscriber pattern.

15. Authentication

• core/auth.py::AuthManager is JSON-file-backed (data/auth.json for users, data/sessions.json for session tokens), not a DB table.
• Two callable-facing auth mechanisms, both resolved in AuthMiddleware (in app.py, gated by AUTH_ENABLED=true):
  • Cookie session auth: bcrypt-password-gated, 7-day TTL, sets request.state.current_user to the resolved username.
  • Bearer API-token auth (Authorization: Bearer ody_...): matched against bcrypt-hashed ApiToken rows with scopes (routes/api_token_routes.py already defines memory:read/memory:write scopes); sets request.state.current_user = "api" plus api_token_owner/api_token_scopes.
  • An internal-tool loopback path lets the agent's own HTTP tool calls reach admin-gated routes via X-Odysseus-Internal-Token / X-Odysseus-Owner.
• Multi-user support is real: AuthManager.users is a dict keyed by lowercase username, with admin-gated create/delete/rename, and reserved usernames that can never be created (internal-tool, api, demo, system).
• Privilege model: DEFAULT_PRIVILEGES includes can_manage_memory among others (can_use_agent, can_use_documents, can_use_bash, etc.). Admins get all privileges unconditionally. Read-only memory routes (GET /api/memory, /timeline) do not require can_manage_memory — only mutation (add/import) does. Per THREAT_MODEL.md, memory management is explicitly listed as available to both admins and non-admins, unlike shell/email/MCP/tokens/settings which are admin-only.
• Ownership pattern to reuse for a graph endpoint: resolve owner = get_current_user(request) (or effective_user(request) for Bearer-token callers), filter the memory load by that owner, and re-verify ownership on any caller-supplied id via the existing _verify_memory_owner helper rather than trusting the id in isolation.

16. Docker configuration

• Two-stage Dockerfile on python:3.14-slim, final image installs build-essential, cmake, curl, git, nodejs, npm, chromium, tmux, openssh-client, gosu, libgl1, ... plus the static Docker CLI (no daemon; host Docker socket is bind-mounted separately if enabled). Exposes port 7000; default CMD is uvicorn app:app --host 0.0.0.0 --port 7000, wrapped by docker/entrypoint.sh.
• docker-compose.yml composes four services:
  • odysseus — the app container; volumes for data, logs, .ssh, .cache/huggingface, .local; huge environment block (LLM endpoints, embeddings, auth flags, upload limits, PUID/PGID); depends_on: searxng (healthy), chromadb (started).
  • chromadb — separate container, image chromadb/chroma:latest, bound 127.0.0.1:8100:8000 on the host, named volume chromadb-data, ANONYMIZED_TELEMETRY=FALSE. No healthcheck (app relies on its own TCP probe + heartbeat() instead).
  • searxng — pinned version, custom entrypoint templating settings.yml, healthcheck via Python urlopen, minimal Linux capabilities (cap_drop: ALL + a small cap_add set).
  • ntfy — push-notification relay, bound 127.0.0.1:8091:80.
• docker/entrypoint.sh implements the standard PUID/PGID drop-privilege pattern (create/reuse group+user matching host PUID/PGID, chown a bounded set of data directories, then exec gosu $ODY_USER "$@" so signals reach uvicorn directly).
• GPU overlays (docker/gpu.nvidia.yml, docker/gpu.amd.yml) and an opt-in docker/host-docker.yml (mounts the Docker socket) are pure compose overlays, not relevant to the Memory Graph View itself but relevant if it ever needs a background job (e.g. embedding recompute) that benefits from GPU access.
• Nothing about the graph feature requires new Docker services: no new database, no new container. It rides on the existing odysseus app container and the existing chromadb container.

17. Testing framework

• Backend: pytest, configured in pyproject.toml (testpaths=["tests"], asyncio_mode="auto", a fixed set of area_* taxonomy markers registered in tests/_taxonomy.py).
• tests/conftest.py is intentionally thin (repo convention: "prefer explicit local setup over hidden global fixtures", per tests/README.md). It forces DATABASE_URL=sqlite:///:memory: before any core.database import, pre-imports modules so later per-file mocking doesn't poison the real ORM for other tests, and stubs a fixed list of optional third-party deps if not installed. There is no shared TestClient/seeded-DB fixture.
• Route-testing convention (seen in tests/test_memory_owner_isolation.py, tests/test_memory_routes_session_owner.py): tests do not spin up a FastAPI TestClient/ASGI app. Instead they call the setup_*_routes(...) factory directly with real or MagicMock() dependencies, look up the target endpoint function off router.routes by path+method, and call it directly with a hand-built Request stand-in (SimpleNamespace(state=SimpleNamespace(current_user=...))). Auth is bypassed by monkeypatching get_current_user/require_user/require_privilege directly on the route module. Ownership tests assert HTTPException(404) on cross-owner access and that returned payloads only contain the caller's own data.
• tests/helpers/ provides sqlite_db.make_temp_sqlite and db_stubs.make_core_db_stub for tests needing a real file-backed SQLite DB. tests/run_focus.py supports marker-based selective runs (e.g. -m area_routes).
• Frontend: no Jest/Vitest/Mocha/Playwright/Cypress. Pure-logic JS files are unit-tested via Node's built-in node:test runner, invoked as a subprocess from a thin pytest wrapper (e.g. tests/test_streaming_segmenter_js.py shells out to node --test tests/streaming/*.test.mjs), or via a vm.createContext() sandbox that string-shims import/export out of a production file before running it headlessly (tests/markdown_codefence_placeholder_regression.mjs). tests/bombadil-spec.ts is a separate fuzzing/property spec (Antithesis-style) for full-app exploration, not part of the normal pytest/CI gate. CI (.github/workflows/ci.yml) runs node --check (syntax only) over all frontend JS, plus pytest -q for everything else.
• Implication for the Memory Graph View: any pure-JS graph-layout/edge-derivation logic should get a .test.mjs file run via node --test, wrapped by a thin pytest shim, following the existing pattern. DOM/rendering behavior remains manually verified against the running app — there is no automated DOM test harness in this repo today.

18. Security posture (relevant to a new data-exposing endpoint)

• THREAT_MODEL.md frames Odysseus as a privileged local-access console for trusted users on a private network, not a public multi-tenant SaaS — but multi-user ownership isolation is still a real, tested boundary (see tests/test_memory_owner_isolation.py).
• core/middleware.py's SecurityHeadersMiddleware already sets X-Content-Type-Options, Referrer-Policy, Permissions-Policy, HSTS (when HTTPS), and a nonce-based CSP (script-src 'self' 'nonce-{nonce}' https://cdn.jsdelivr.net) on all normal routes. A new graph view page doesn't need a special CSP branch unless it renders inline HTML/iframe content the way the visual-report/tool-render pages do.
• CORS is configured via CORSMiddleware with allow_origins from ALLOWED_ORIGINS (default http://localhost,http://127.0.0.1) and allow_credentials=True.
• Known gaps documented in THREAT_MODEL.md (no shell/filesystem sandbox for agent tools, an SSRF gap in chat base_url, coarse API token scopes) are not directly implicated by a read-only graph endpoint, but the "coarse token scopes" gap matters if the graph is ever exposed over a Bearer API token — it should reuse the existing memory:read scope rather than inventing a new one.

19. Safest extension points, summarized

Ranked from safest/most isolated to most invasive:

1. New read-only backend route GET /api/memory/graph inside the existing routes/memory/memory_routes.py module (or a new routes/memory/memory_graph_routes.py included the same way), reusing _owner(request) scoping and the existing MemoryManager/MemoryVectorStore singletons already wired up in app_initializer.py. No new tables, no new services, no new containers.
2. New frontend module static/js/memoryGraph.js, registered with Modals.register(...) the same way memory-modal already is, either as a new tab inside the existing Brain modal or a sibling modal reachable from it. No changes to app.js's routing map are strictly required if it's a tab; one new _routeOpen entry if it's a standalone deep-linkable view.
3. A vendored graph-rendering library dropped into static/lib/, following the existing vendoring convention (no npm/bundler involvement).
4. Optional: a new SSE channel for live updates, modeled directly on src/agent_runs.py's per-subscriber asyncio.Queue pattern, with new memory_updated/memory_deleted fire_event() calls added at the existing pin/update/delete call sites. This is additive and does not touch the existing task-automation consumer of memory_added.
5. Not recommended as a first step: touching the dormant SQL memories table, since it is not the runtime source of truth today and reconciling it would be a separate, larger migration unrelated to shipping a graph view.

No area inspected requires a new database engine, a schema migration to an existing hot-path table, or a new Docker service.