https://engraphis.com/

https://discord.com/invite/Wfr2ejBmY

Give your AI agents a memory. See it, search it, and watch it self-maintain — all in a beautiful WebUI on your own machine.

^{Knowledge Graph · run engraphis-dashboard to see it live}

Open-Source users: Remember to Update regularly! Improvements and fixes twice a day. Invite your friends!

Beta: the Team layer (multi-user dashboard, seats, roles, audit log, team invite emails, cloud sync relay) is early-access beta — expect rough edges and breaking changes before it stabilizes. The single-user engine, dashboard, and MCP server are stable.

pip install "engraphis[server]"
engraphis-dashboard

Opens http://127.0.0.1:8700 in your browser. No cloud, no signup, no API key for memory. Memory lives in a local SQLite file on your machine. When hosted user accounts are enabled, their credentials and sessions live in a companion <database>.users.db; back up both files.

You'll see the full product — a dark-themed (with multiple theme options in left sidebar), sidebar-navigated dashboard with 14 tabs:

New graphical interface! Shape the Knowledge Graph with several Styles, Colors, and Presets. Switch among Cyberpunk, Galaxy, Solar system, and Classic looks; choose a color palette and layout preset; or change the colors used for each type of node.

The dashboard is powered by the v2 engine — the same MemoryService that backs the MCP server and the Python library. What you see in the UI is what your agents get.

The dashboard has the focused memory-inspection view built in — no separate app or port:

• Open any memory to see its supersession chain with word-level diffs — exactly when a fact changed and why
• Offline knowledge graph (vendored renderer — no CDN, works air-gapped)
• Score breakdowns on every recall, Why/Timeline/link browsing, proactive recall, consolidation, audit trail
• Keyboard-navigable, ARIA-annotated, light/dark mode

The standalone Inspector (:8710) was retired 2026-07-10 and folded into the one dashboard on :8700.

Your agents forget everything between sessions. Engraphis fixes that — on your machine. Every new session, your coding agent starts from zero: re-asking which package manager you use, re-learning the codebase, forgetting why you chose PASETO over JWT. Engraphis gives agents durable, scoped, explainable memory.

Under the hood: Ebbinghaus forgetting-curve decay, interaction-aware reinforcement, bi-temporal facts, and hybrid (vector + lexical + graph) recall. The engine is 100% local: SQLite + local embeddings. You bring an LLM only for optional chat, synthesis, structured extraction, or structured consolidation.

• Local-first & private — runs offline; the core depends only on numpy.
• MCP-native — 28 tools for Claude Code, Command Code, Cursor, Cline, Zed, Windsurf.
• Self-maintaining facts — writes are deterministically conflict-resolved (no LLM required).
• Advisory retention supervision — an optional LLM can label writes as ephemeral, normal, or critical; outputs are bounded, clamped, audited, and can never silently drop a write.
• Principled recall — six-term score over retention, semantic, lexical, graph, importance, recency.
• Bi-temporal truth — contradictions invalidate instead of overwriting (engraphis_why / engraphis_timeline).
• Grounded, not guessed — cited answers or explicit abstain; provenance on every memory.
• Task-ready context — bounded proactive packets combine task/agent state, cited memories, suggested follow-ups, and the last-session handoff; optional LLM prose is accepted only when its citations validate.
• Composable intelligence — opt-in deterministic conflict triage (duplicate / refinement / contradiction / obsolete) and UserModel recall reranking helpers; neither changes default recall unless called.
• Human-governed lifecycle — pin, forget, correct, promote to a wider scope, and manually merge several memories into one without deleting their history; every change is audited.
• One layered graph — temporal, entity, causal, and semantic overlays share the same database, with persistent code↔memory links and intent-aware recall.
• Privacy-safe receipts — remember, link, recall, and indexing operations can be verified through a content-free SHA-256 receipt chain without exporting memory or query text.
• Code-aware — incremental multi-language symbol/call/import graph, code↔memory links, path queries, communities/hotspots, git/PR impact analysis, and portable graph exports.
• Sleep-time consolidation — scheduled job distills recurring episodes, reports its compaction.
• Scoped — workspace → repo → session hierarchy.
• Encryption at rest — optional SQLCipher (AES-256) whole-database encryption via ENGRAPHIS_DB_KEY. No plaintext fallback when a key is set.
• Cloud sync — cross-device and cross-team memory sync with deterministic CRDT merge (folder transport for self-hosting, managed relay for zero-setup). One-click "Sync now" or automatic cadence in the dashboard.
• Import & ingest — local documents/code/DOCX plus optional PDF OCR, image OCR, audio/video transcription, and live PostgreSQL schema introspection.

The dashboard ships as a Docker image that defaults to the v2 dashboard (multi-user auth, roles, seats, cloud-license revocation). Deploy one instance on Railway and access your memories from any browser. Two supported paths:

• Pro solo — a Pro member deploys a single-admin cloud instance: browser dashboard (analytics, automation, export) + a self-hosted sync relay. Activate the same Pro key on each local instance, set ENGRAPHIS_RELAY_URL on both the hosted service and local instances to your Railway deployment URL, then enable auto-sync (or run Sync now). Keep ENGRAPHIS_CLOUD_URL=https://team.engraphis.com on the hosted service so trials, revocable leases, and fallback invite delivery continue to use the managed issuer. One admin, no member seats.
• Team admin — a Team administrator deploys one instance and invites members (email + password + role). Members sign in at your URL and connect their agents over HTTP/MCP — no local install for members.

See docs/HOSTING_RAILWAY.md for the 5-minute guide covering both paths (volume, custom domain, activate Pro/Team, create the first admin, invite members, and connect agents).

→ Deploy on Railway (5-minute guide)

Deploying from this repo's Dockerfile takes about five minutes: create the service, attach a persistent /data volume (so activated keys + memories survive redeploys), and set ENGRAPHIS_FORWARDED_ALLOW_IPS=*, ENGRAPHIS_DASHBOARD_URL, and a strong ENGRAPHIS_API_TOKEN for hosted bootstrap. railway.json in this repo already supplies the build and healthcheck config. The hosting guide walks through each step with the exact values.

(A one-click "Deploy on Railway" button previously sat here pointing at railway.app/new?template=<raw railway.json URL>. Railway ignores that parameter — railway.json is per-service build config, not a publishable template — so the button only ever landed people on a generic project chooser. It was removed on 2026-07-18 rather than left looking functional. docs/RAILWAY_TEMPLATE.md contains the spec to publish a real template; once published, its button can go back here.)

Hosted Agent Connect tokens are per-user, shown only once, and stored only as SHA-256 digests. Roles are rechecked on every HTTP/MCP call; disabling a member or resetting their password permanently revokes existing agent tokens. The hosted /mcp endpoint exposes the same 28-tool service as local engraphis-mcp. See the Agent Connect guide.

pip install "engraphis[all]"        # dashboard + MCP server + code graph + available platform extras
pip install "engraphis[server]"     # dashboard + REST API
pip install "engraphis[mcp]"        # MCP server only
pip install "engraphis[documents]"  # PDF + image OCR bindings
pip install "engraphis[transcription]" # faster-whisper audio/video
pip install "engraphis[postgres]"   # PostgreSQL schema introspection
pip install "engraphis[encryption]" # SQLCipher encryption-at-rest extra
pip install engraphis               # core library — numpy only, fully offline

The NumPy-only core library supports Python 3.9+. Current patched releases of the WebUI stack, MCP SDK, and image parser require Python 3.10+, so use Python 3.10 or newer for the server, mcp, documents, or all installation paths. sqlcipher3-binary publishes CPython manylinux x86-64 wheels. On that target, engraphis[encryption] installs the driver. The cross-platform all extra deliberately omits it so all remains resolvable on macOS, Windows, Linux ARM, and musl; on those targets, provision a compatible SQLCipher driver separately before enabling a database key. Plaintext SQLite remains the explicit default on every platform.

Linux / macOS: if pip install fails with error: externally-managed-environment, your system Python is marked read-only (PEP 668). Install into a virtual environment instead — python3 -m venv venv && source venv/bin/activate && pip install "engraphis[server]" — or use Docker (docker compose up). pipx install "engraphis[server]" also works.

First run downloads all-MiniLM-L6-v2 (~80 MB). Without it, the engine falls back to a deterministic offline embedder so it always runs.

pip install "engraphis[server]"
engraphis-dashboard                   # → http://127.0.0.1:8700
engraphis-dashboard --install-shortcuts   # → Desktop + Start Menu icons

docker compose up                     # → http://127.0.0.1:8700

A fresh clone needs no .env: the default service runs engraphis-dashboard --no-open, stores the v2 database plus license state on a named volume mounted at /data, and accepts overrides from .env or the shell. The legacy v1 API is opt-in with docker compose --profile api up engraphis-api and uses a separate database so its incompatible schema cannot collide with the dashboard.

Compose publishes both services on host loopback only. Set a strong ENGRAPHIS_API_TOKEN before changing either port mapping to a non-loopback host address.

Set ENGRAPHIS_API_TOKEN to require API authentication, ENGRAPHIS_DB_KEY to encrypt the database at rest, and ENGRAPHIS_LICENSE_KEY to unlock Pro/Team features. See docker-compose.yml for all options.

pip install "engraphis[mcp]"
engraphis-init                     # writes .env + prints config snippets
claude mcp add engraphis -- engraphis-mcp
cmd mcp add engraphis -- engraphis-mcp  # Command Code CLI

Your agent now has 28 tools — remember, recall (grounded + proactive), proactive context, grounded answer alias, why, timeline, forget, pin, correct, promote, ingest, consolidate, index_repo, search/code path/impact/export, privacy receipts, PostgreSQL schema ingestion, link, record_event, start/end_session, and stats. See the MCP tools table below.

For unattended jobs, engraphis_start_session, engraphis_remember, and engraphis_record_event use workspace default when workspace is omitted.

pip install "engraphis[code]"
engraphis-graph index -w acme -r api --root .
engraphis-graph search -w acme -r api "UserService"
# `query`/`explain` blend code search with your stored memories: query matches symbol
# and file NAMES (a full question sentence won't match anything), and explain's answer
# is drawn from memories recorded against the repo — both are empty on a fresh index.
engraphis-graph query -w acme -r api "UserService"
engraphis-graph explain -w acme -r api "why does deploy depend on approval?"
engraphis-graph path -w acme -r api UserService DatabasePool
engraphis-graph impact -w acme -r api --root . --git-range origin/main...HEAD
engraphis-graph prs -w acme -r api --base main --head HEAD
engraphis-graph export -w acme -r api -o engraphis-graph-out
engraphis-graph install-merge-driver --root .

The export contains graph.json, a self-contained graph.html, and GRAPH_REPORT.md. Indexing supports Python, JavaScript, TypeScript, Go, Rust, Java, C#, C, C++, SQL, and Terraform. Tree-sitter is used when available; the dependency-free regex backend remains a functional fallback. Definitions, methods, calls, imports, ownership, variables, inheritance/implementation, and docstrings/comments are indexed. Indexing is incremental by content hash, honors .engraphisignore, and does not follow file symlinks outside the repository root. Call edges are name-based and best-effort rather than type-resolved. The optional Git merge driver validates bounded graph JSON and deterministically unions nodes and edges instead of choosing one export side.

For a read-only recall and graph API that can be shared without exposing write operations:

pip install "engraphis[server]"
engraphis-graph-server                 # API at http://127.0.0.1:8720; schema at /openapi.json

A non-loopback bind fails closed unless ENGRAPHIS_GRAPH_TOKEN (or ENGRAPHIS_API_TOKEN) is set. See the v3 architecture/design document.

from engraphis.service import MemoryService

mem = MemoryService.create("engraphis.db")
mem.remember("Auth migrated from JWT to PASETO.", workspace="acme", repo="api")
hit = mem.recall("why did we change auth?", workspace="acme", repo="api")
print(hit["context"])

The same MemoryService backs the dashboard and the MCP server.

Engraphis separates automatic write resolution from explicit human governance:

Manual N→1 merge is available through MemoryService.merge() and POST /api/merge:

a = mem.remember("Deploys happen Friday at 3pm.", workspace="acme")
b = mem.remember("We deploy Fridays around 15:00.", workspace="acme")

merged = mem.merge(
    [a["id"], b["id"]],
    "Deploys ship every Friday at approximately 15:00.",
    workspace="acme",
    reason="deduplicate the deployment schedule",
)
print(merged["compaction"])

All sources must belong to the named workspace. The result inherits the strictest source sensitivity, remains untrusted if any source was untrusted, and stays pinned if any source was pinned. The full multi-predecessor chain remains visible through inspection, Why, and Timeline.

The core engine, single-user dashboard, standalone MCP server, and governance tools are free and Apache-2.0, permanently. Paid Pro/Team keys are server-authoritative: the vendor signature is checked locally, then the key must hold a current machine-bound lease from the configured/vendor relay. Revoked, expired, or seat-exceeded keys fail closed; an unexpired lease provides bounded grace for transient network failures. Pro is $10/mo ($100/yr), Team is $20/seat/mo ($200/seat/yr), and the dashboard offers a 3-day server-issued Pro or Team trial after email confirmation — no card required.

Team is early-access beta. Multi-user logins, seats, roles, the team audit log, team invite emails, and the cloud-sync relay are all in active development — expect rough edges and breaking changes. Pro (single-user paid features) is stable. Free is stable.

Memory relationships, extracted entities, and code structure stay normalized in one SQLite database. Edges are tagged as temporal, entity, causal, or semantic, so callers can select a logical overlay without maintaining separate graphs. Schema-v3 migration is additive and idempotent: existing memories and bi-temporal history remain in place, while legacy edge layers are inferred once.

MemoryService.intent_remember(), intent_link(), and intent_recall() provide a transport-neutral agent protocol while the existing engraphis_remember, engraphis_link, and engraphis_recall tools remain the canonical MCP vocabulary. Explicit links can persist both a layer and a durable rationale. Intent recall maps explain, summarize_history, and locate_code to appropriate layer filters; code intents also return matching symbols when a repository is supplied.

The operation-receipt chain is deliberately content-free. It records bounded operation metadata and chained hashes, while excluding raw memory/query text, workspace names, memory IDs, and actor identities from exported receipt payloads. Use engraphis_receipts, engraphis_verify_receipts, and engraphis_export_receipts to inspect the chain or compare it with a previously saved head/count anchor. A separately maintained local count/head anchor and persistent integrity marker make interior edits, reordering, and tail truncation detectable.

See the v3 architecture document for the data flow and SECURITY.md for the trust boundaries.

Cloud sync keeps your memory store consistent across all your machines — and, on the Team tier, across a group — without giving up local-first ownership. It ships two transports:

• Folder transport — any shared directory (Dropbox, iCloud, Syncthing, a git repo, a mounted drive). Zero infrastructure.
• Managed relay — HTTPS against the Engraphis relay, authenticated by your license key. One-click in the dashboard or python -m scripts.sync --relay.

Sync is a state-based CRDT: deterministic merge, no conflict copies, no data loss. Every field resolves by a commutative, idempotent rule so merge(A, B) == merge(B, A). The current sync format carries memories and their memory-to-memory links; entity/code graph reconciliation is not yet part of sync. secret memories are never exported, and Team personal folders are never uploaded to a shared-account relay. Managed-relay traffic uses HTTPS, but bundles are not yet client-side end-to-end encrypted or zero-knowledge. See docs/SYNC.md for architecture, security model, and CLI usage.

The current shared and commercial surfaces enforce:

• Team authentication and RBAC — first-admin setup is atomic; login PBKDF2 verification runs outside the shared store lock; sessions and agent-token history are bounded; disable/reset events revoke long-lived tokens. Viewers read and members perform ordinary writes. Only admins can change account-wide sync policy, import/index server-side resources, or delete/merge workspaces.
• License and billing lifecycle — paid features require a current machine-bound lease; process cache and device-id creation are serialized. Billing webhook fulfillment is bounded, durable, retry-safe, and idempotent.
• SQLite transaction safety — shared v2 connections serialize complete write transactions; a failed statement that opened a transaction rolls it back and releases its lock. Legacy decay is frequency-independent, and sync preserves future bi-temporal validity horizons.
• Relay isolation — workspace allow-lists are enforced while applying fetched data, personal Team folders cannot enter the shared-account relay, and device-local secret memories cannot be remotely overwritten, invalidated, or downgraded. Bundle size, count, and per-workspace storage are bounded.
• Hostile-input handling — sync-folder peers, graph merge inputs, repository walks, resource files, and PostgreSQL selectors are treated as untrusted; traversal, symlink/replace races, oversized/deep payloads, malformed rows, and non-finite JSON are rejected.
• Proxy and network hardening — default loopback CORS follows ENGRAPHIS_PORT; proxy-reported HTTPS produces Secure session cookies, and redirects use the configured dashboard URL rather than a caller-controlled Host header. Managed-service clients reject insecure or malformed endpoints and never forward bearer credentials across HTTP redirects.

See SECURITY.md for supported versions, deployment requirements, known gaps, and the vulnerability-reporting process.

Set ENGRAPHIS_DB_KEY (or ENGRAPHIS_DB_KEY_FILE) and install the extra:

pip install "engraphis[encryption]"

The entire database file is transparently encrypted with AES-256 via SQLCipher — full-text search, the graph, and every query keep working unchanged. When a key is set, Engraphis fails loud rather than silently falling back to plaintext. Generate a strong key:

python -c "import secrets; print(secrets.token_hex(32))"

An existing plaintext database cannot be opened with a key — migrate it (dump → import into a fresh keyed DB). See .env.example for all encryption options.

Drag-and-drop or server-side import, role-gated and bounded:

• Dashboard upload — accepts text, Markdown, code, JSON/CSV/HTML, DOCX, and exported Google Workspace documents directly; optional adapters add PDF text extraction, image OCR, and audio/video transcription. Native .gdoc pointer files contain no document body, so export them as DOCX, PDF, HTML, or plain text before local ingestion.
• Server-side folder import — MemoryService.import_folder() reads a directory on the machine running Engraphis. Large resources are chunked deterministically even when the configured extractor is none; path-traversal guards still apply.
• PostgreSQL — engraphis_ingest_postgres_schema, POST /api/resources/postgres, or engraphis-graph postgres converts tables, columns, constraints, and foreign keys into a schema memory and entity graph. The DSN is never persisted.
• MCP ingest — engraphis_ingest accepts raw text and applies the configured extractor (chunk, llm, or llm_structured); with none it stores one verbatim memory.
• Sub-file chunking — set ENGRAPHIS_EXTRACTOR=chunk to split long, multi-topic documents into retrieval-sized, structure-aware pieces (headings start new chunks; ~256-token target with sentence-level overlap) without an LLM. Each chunk becomes its own memory, so recall returns the relevant passage instead of a whole file — a big context-reduction win on long docs. Works across all three ingest paths (dashboard upload, import_folder, and engraphis_ingest). Measure the payoff with the bundled eval: python -m eval.chunking_eval --dataset eval/datasets/longdoc.jsonl --k 5 (whole-file vs. chunked, same recall pipeline, offline).
• Structured LLM extraction — ENGRAPHIS_EXTRACTOR=llm_structured validates typed facts, entities, relations, keywords, and confidence before storage. Its preserved entity/relation metadata feeds the knowledge graph automatically.

Files imported through the dashboard or import_folder() are marked untrusted by default; MCP ingest remains an authenticated agent write.

Manual consolidation is free. The Pro Automation tab (and the GET/POST /api/automation plus POST /api/maintenance/run API) can keep the store clean without you clicking anything, using a maintenance policy with two modes that compose:

• Scheduled maintenance — a consolidation + retention sweep on a fixed cadence (cadence_hours). Recurring episodic memories are distilled into semantic digests, and memories fading below archive_below retention are archived bi-temporally (pinned memories are always protected).
• Auto-dreaming — a background consolidation + cross-cluster inference loop (no cron needed — it runs inside the dashboard process) that fires when both hold: the store has accumulated ≥ dream_min_new new episodic memories since the last sweep, and the store has been idle for dream_idle_minutes. Dreaming emits low-salience dream_inference memories (cross-cluster/entity profiles, marked untrusted and linked back to their sources) so inferred knowledge is auditable and never silently promoted.

Knobs (dashboard Automation tab ↔ /api/automation API): enabled, cadence_hours, consolidate, min_cluster, archive_below, dream, dream_min_new, dream_idle_minutes, infer. Headless / no-dashboard-open: python -m scripts.auto_maintain --apply (via Task Scheduler or cron).

Manual consolidation can also use schema-validated LLM output through MemoryService.consolidate, POST /api/consolidate, engraphis_consolidate, or python -m scripts.consolidate --structured. Source memories remain live by default; supersede_sources / --supersede-sources closes them only after validated replacement facts are written.

All via environment (or .env):

See .env.example for the full list including commercial/vendor, email delivery, and cloud-license enforcement options.

engraphis/
├── engraphis/
│   ├── core/                # v2 engine — interfaces, store, recall, scoring, schema, sync
│   ├── backends/            # pluggable embedder / vector index / reranker / codegraph / sync transports / encryption
│   ├── service.py           # validated MemoryService facade
│   ├── mcp_server.py        # MCP server — 28 tools
│   ├── dashboard_app.py     # dashboard WebUI (FastAPI)
│   ├── read_only_api.py     # token-protected recall/repository-graph HTTP surface
│   ├── autosync.py          # background auto-sync loop (Pro/Team)
│   ├── licensing.py         # signed-key + live machine-bound lease verification
│   ├── analytics.py         # Pro analytics engine
│   ├── automation.py        # scheduled maintenance policies (Pro)
│   ├── billing.py           # Polar webhook fulfillment
│   ├── config.py / app.py   # env settings / REST server
│   └── static/              # dashboard frontend
├── eval/                    # offline retrieval eval harness + datasets
├── tests/                   # pytest suite (300+ tests, offline numpy-only core)
├── scripts/                 # start_dashboard, inspector, cli, init, consolidate, sync
├── docs/                    # SYNC.md, KILO_CODE_INTEGRATION.md
├── Dockerfile / docker-compose.yml
└── pyproject.toml

New capability belongs in the v2 path (engraphis/core/, engraphis/backends/, and MemoryService) behind the interfaces in core/interfaces.py. The flat-namespace v1 server under engraphis/app.py, routes/, stores/, and engines/ remains a compatibility/reference surface; engraphis-dashboard, the MCP server, and the Python quickstart above use v2.

The offline quality gate (no network, no API key):

pip install numpy pytest ruff
python -m pytest tests/ -q
python -m eval.harness --dataset eval/datasets/sample.jsonl --k 5
python -m eval.harness --dataset eval/datasets/codemem.jsonl --k 5
python -m eval.ablation
ruff check .

Numbers, not assertions: the offline harness is a correctness floor (deterministic embedder). LoCoMo / LongMemEval competitive numbers run separately with a real embedder — see BENCHMARKS.md.

Apache-2.0 — see LICENSE and NOTICE. "Engraphis" is a trademark of the Engraphis project; the license does not grant trademark rights.