Memori-Vault

English: current page
中文文档: README.zh-CN.md
Contributing: CONTRIBUTING.md | 中文贡献指南 Tutorial: docs/TUTORIAL.md | 中文辅助: docs/TUTORIAL.zh-CN.md

Memori-Vault is a local-first memory engine for personal and team knowledge. It combines semantic chunking, vector retrieval, and asynchronous Graph-RAG extraction on Ollama + SQLite, while keeping first-answer speed stable under background indexing.

Highlights (Current)

• Local-first ingestion pipeline (.md / .txt) with watcher + semantic chunking.
• Structured retrieval pipeline with document routing, chunk retrieval, citations, and evidence output.
• Async indexing refactor:
  • fast path for searchable chunks first
  • deferred graph build in background queue
• Indexing strategy controls:
  • continuous | manual | scheduled
  • resource budget low | balanced | fast
  • pause/resume + trigger reindex
• Settings center (right drawer):
  • UI language and AI answer language (separate)
  • model provider profiles (local Ollama / remote OpenAI-compatible)
  • watch folder switching
  • top-k retrieval control
  • personalization (font, size, theme)
• Search scope selector:
  • nested folder expand/collapse
  • multi-select files/folders
• Source cards:
  • markdown preview for .md
  • expand/collapse
  • open file location

Current Validation Status

• Local-first runtime and enterprise policy gates are implemented.
• Citation validity is currently strong in the checked-in offline regression corpus.
• Retrieval precision is not yet at a strong mixed-corpus bar.
  • core_docs offline baseline: Top-1=0.6970 on 6 indexed documents
  • repo_mixed offline baseline: Top-1=0.4773 on 11 indexed documents
• These are small checked-in regression baselines, not a validated 50k-document accuracy result.
• Live local-model validation is still blocked on local Ollama / embedding availability on the current machine.

Current posture:

• docs-only retrieval is usable as an internal baseline
• mixed-corpus retrieval should still be treated as beta/internal validation, not as a finished accuracy claim

Details: docs/RETRIEVAL_BASELINE.md

Runtime Modes

1. Desktop mode:

• Tauri shell + UI + IPC backend.

2. Server mode:

• memori-server exposes HTTP APIs for local/browser access and private deployment.
• The current product experience is desktop-first; browser-facing UI support is still being aligned with the server runtime.

Enterprise (Private Deployment v1 Preview)

• Single-tenant private deployment for engineering organizations.
• Preview auth/session entry plus API RBAC (viewer/user/operator/admin).
• Admin APIs for health, metrics, policy, audit, reindex, pause/resume.
• Model governance: local-first with remote egress allowlist enforcement.
• Deployment assets included (deploy/systemd, env template, backup/restore scripts).

Current note:

• Enterprise deployment is available as a private deployment preview in v0.3.0.
• Auth/session flows are suitable for controlled internal environments first and will continue to harden in later releases.

Details: docs/enterprise.md

Architecture

Workspace crates:

• memori-vault: watch/debounce/event stream
• memori-parser: parse/chunk
• memori-storage: SQLite + vector/graph/task metadata
• memori-core: orchestration, retrieval, indexing worker
• memori-desktop: Tauri commands and desktop lifecycle
• memori-server: Axum APIs
• ui: React + Vite + Tailwind v4 frontend

Development Quick Start

cargo fmt --all -- --check
cargo clippy --workspace -- -D warnings
cargo test --workspace
pnpm --dir ui run build

Desktop dev:

pnpm --dir ui run dev -- --host 127.0.0.1 --port 1420 --strictPort
cargo tauri dev -p memori-desktop

Server dev:

cargo run -p memori-server
pnpm --dir ui run dev -- --host 127.0.0.1 --port 1420 --strictPort

Notes

• Ollama local runtime is recommended for local provider mode.
• Remote provider mode is optional and user-configured.
• Enterprise policy can enforce local_only or remote allowlist mode.
• Legacy theme key memori-theme-mode is migration-only; active key is memori-theme.

License

Apache License 2.0.