DevMemo AI

Chinese

DevMemo AI is a self-hosted developer knowledge base for individuals and small teams. It is built on Memos, which remains the source of truth for original Memo data, user identities, and visibility permissions. A separate FastAPI AI Service stores AI-derived state only; it does not introduce a second identity or authorization system.

Use DevMemo AI to capture code snippets, bug reports, solutions, and technical decisions as traceable developer knowledge. After human review, accepted insights can be compiled into bounded, copyable Context Packs for an IDE or a subsequent tool.

Unofficial downstream project. DevMemo AI is not affiliated with, endorsed by, or supported by the Memos project. Read UPSTREAM.md and NOTICE for the upstream baseline, ownership boundary, and attribution.

Key capabilities

• Keep Memo data, users, and permissions in Memos; AI state never replaces the Memos source of truth.
• Generate structured templates and AI-derived insights for Code Snippets and Bug Reports.
• Review and accept or reject insights in the Memo detail view before they can contribute to a working context.
• Build context-pack-v1 from explicitly visible Memos and accepted insights. It contains only safe titles, summaries, and source references—not raw Memo content, Webhook payloads, secrets, or chunk content.
• Run locally with deterministic providers and in-memory retrieval by default, with a low-resource, offline-first baseline.
• Keep FastEmbed, Qdrant, Ollama, Webhook indexing, and public chunk retrieval as explicit opt-ins rather than default behavior.
• Offer an experimental, read-only Evidence Answer entry through the Memos BFF. It is disabled by default and returns only bounded answers, server-owned citations, and a redacted execution trace.
• Run a fixed project_summary AgentRun through a default-disabled, authenticated same-origin BFF. Memos resolves one to ten visible Memo revisions before AI Service synchronously executes the bounded deterministic plan and returns a creator-bound Markdown artifact.

Current project scope

The local personal-demo loop is complete: capture project notes in a Memo, run the fixed project_summary AgentRun, observe the bounded execution state, and preview or download the generated Markdown artifact. The flow is authenticated, visibility-aware, deterministic by default, and does not require an external model.

DevMemo AI is not presented as a general autonomous Agent or a production-ready multi-instance AI platform. Background workers, approval flows, Memo write-back, free-form Agent tasks, real-user external-Provider acceptance, and shared multi-instance runtime state remain outside the current product boundary.

Architecture and data boundaries

Memos (Go + React)
  ├─ Memo data, users, and permissions: source of truth
  └─ explicit Webhook integration
             │
             ▼
FastAPI AI Service
  ├─ AI-derived SQLite state only
  ├─ templates, insights, and optional index state
  └─ provider/vector-store adapters
             │
             ▼
Memo detail view
  ├─ AI insights review
  ├─ in-memory Context Pack generation and copy
  └─ bounded project-summary AgentRun and Markdown artifact

See docs/structure.md for repository and runtime boundaries, and docs/api.md for API contracts. The experimental Agent design and remaining delivery gates are documented separately in docs/agent-architecture.md and docs/agent-development-roadmap.md. The sanitized evaluation method/results and current completion gates are recorded in docs/agent-evaluation-benchmark.md and docs/r6-completion-audit.md.

Quick start

Prerequisite: Docker Desktop with Docker Compose.

git clone https://github.com/ToYOhin/devmemo-ai.git
Set-Location devmemo-ai
docker compose config
docker compose up -d --build

After startup:

• Memos: http://localhost:5230
• AI Service health: http://localhost:8000/health

The published stable image is also available directly:

docker pull ghcr.io/toyohin/devmemo-ai:stable

The stable image publishes linux/amd64, linux/arm64, and linux/arm/v7 manifests. Download native executables from GitHub Releases.

Run the local Agent demo

The source-build demo requires Docker Desktop, Node.js 24 or later, and pnpm 11. From the repository root, run:

.\scripts\start-agent-demo.ps1

Then open http://localhost:5230, sign in, create a Memo containing project notes, open its detail view, and choose Build draft in the Project summary panel. The run executes synchronously and shows a Markdown preview with a download action.

.\scripts\start-agent-demo.ps1 -Action status
.\scripts\start-agent-demo.ps1 -Action stop

The launcher generates a process-local delegation secret, keeps the Provider deterministic, does not publish the AI Service port, restores the managed shell environment after completion, and preserves Docker volumes when stopped.

Default security and resource posture

The default configuration is intentionally lightweight and conservative:

AI_INDEX_ON_WEBHOOK=false
AI_INDEX_MODE=memo
AI_VECTOR_STORE=memory
AI_PUBLIC_CHUNK_RETRIEVAL=false
AI_AGENT_ENABLED=false

• Default Compose does not allow private-network Webhook targets.
• The default stack runs only Memos and the AI Service with modest CPU budgets.
• Qdrant and Ollama require explicit Compose profiles.
• Public chunk retrieval remains disabled. It requires a real trusted gateway, Memos visibility mapping, and a verified disable-and-rollback path.
• The Evidence Answer Agent remains an opt-in experiment. The reviewed R6 baseline is published at v0.2.0; later R7 slices add frozen AgentRun contracts, single-host derived-only SQLite persistence, a bounded runtime, and authenticated execution/artifact BFF routes for the fixed project_summary task kind. The Memo detail UI can preview or download the Markdown artifact; a worker, approval flow, Memo write, and multi-instance operation remain separate slices.
• A bounded DeepSeek adapter is available only through explicit configuration. Deterministic remains the default; the external endpoint smoke uses synthetic evidence and does not establish real-Memo or production deployment readiness.

For a controlled local Docker development topology where a Memos Webhook must target ai-service, use the docker-compose.local-webhook.yml override documented in README_AI.md. Do not use that override for a public or multi-user deployment.

Further reading

README_AI.md documents AI Service configuration and optional adapters. docs/operations.md covers deployment, backup, restore, and upgrades. Contributions are welcome through CONTRIBUTING.md.

Support, security, and governance

• Setup and usage support: SUPPORT.md.
• Bugs and feature requests: use this repository's GitHub issue forms.
• Security reports: follow SECURITY.md; do not disclose vulnerabilities in public issues.
• Governance, maintainer responsibilities, and release expectations: GOVERNANCE.md.
• For Memos behavior unchanged by this project, use the upstream Memos support and issue channels.

License

DevMemo AI is distributed under the MIT License. It contains downstream changes based on Memos; upstream attribution and licensing details are preserved in NOTICE and UPSTREAM.md.