---

On Monday your agent learns that payment writes need idempotency keys. On Tuesday, in a new session or a different editor, it starts from zero. AI-DLC Engineering OS gives every MCP-capable agent on your machine the same persistent project memory, plus 26 specialist contracts, and nothing becomes a rule without evidence and your approval.

The memory server runs locally with SQLite: no account, telemetry or hosted service. Your AI client controls which retrieved context it sends to its model provider.

Quick start

Install uv, then run this single command on Windows, macOS or Linux:

uvx ai-dlc-engineering-os setup

Choose Codex, Claude Code or VS Code, review the proposed changes, then confirm. Setup installs a persistent runtime, backs up existing configuration and memory, registers MCP and verifies the selected editors. Restart your editor afterward. VS Code must have been opened once to create its default profile.

Then just work. Open your project and ask:

What do we already know about this codebase before I change the payment flow?

For Claude Code and Codex, setup adds instructions to CLAUDE.md and AGENTS.md that tell the agent to load the project's decisions, lessons and open incidents before significant work. When you decide something worth keeping, say so ("remember that we use idempotency keys on payment writes") and it is available in the next session, in any configured editor.

Add --editors codex,claude --dry-run to preview without applying changes; for unattended setup choose --editors explicitly and add --yes. The installation guide covers requirements, updates, pinned versions, backups and uninstall.

Why not just CLAUDE.md or AGENTS.md?

Keep them. They are the right place for instructions you write by hand. This adds what a static file cannot do well:

	Instruction file	AI-DLC Engineering OS
Who writes it	You, by hand	Agents propose; you approve rules
Shared across editors	One file per tool	One local database for every configured editor
Grows without bloating context	Whole file loaded every time	Only what is relevant to the task is retrieved
Stale knowledge	Edited silently	Superseded with history; the repository and tests win over memory
Secrets pasted by accident	Saved as written	Refused on write

Proof, not promises

Every number below comes from a command in this repository, so you can rerun it.

Check	Result	Reproduce
Cross-session recall (animation above)	Decision recalled by a new process; credential refused and absent from context	python scripts/demo_cross_session.py
Swarm lease reliability, 500 tasks, 8 workers	500/500 completed, 0 duplicates, 0 lost	python benchmarks/run_reliability.py --tasks 500 --workers 8
Retrieval over 1,000 memories	200/200 searches returned results, 104 ms mean	ai-dlc benchmark --operations 1000 --searches 200
Test suite	94 tests, including a real MCP stdio session	pytest

Measured on a 2-vCPU Linux container with Python 3.13. These are reference numbers for one machine, not capacity claims; see the benchmark protocol. CI keeps the demo and both benchmarks runnable.

What you get

Capability	What it does
Memory that stays true	FTS5/BM25 + local vector search over decisions, lessons, incidents and outcomes. Superseding instead of silent overwrite. The repository and its tests always win over memory.
Safe by default	Credentials are refused on every write. Instruction-like content is flagged and withheld from context. Hard purge for data that should not be there.
Human governance	Agents propose; only a human approves rules (ai-dlc-admin, interactive terminal only).
26 specialists	Each with a contract, procedure, checklist, output format and escalation rules.
Measured agent evolution	Golden cases, quality/safety/non-regression gates, shadow → canary → active, hash-chained receipts, drift detection.

How it works

A typical session, driven by your agent through its MCP tools:

1. resolve_project_id → get_project_context(project_id, task): relevant decisions, lessons and open incidents.
2. list_agents → get_agent_context("principal-architect", project_id): the right specialist, with the project's approved rules.
3. Do the work. Persist only what is durable: save_decision, record_learning, record_incident.
4. record_task_outcome with evidence. Learnings stay candidates until verified; rules need human approval.

Memory lives in a single SQLite file (~/.ai-dlc/memory/memory.db, WAL mode), shared by every host on the machine. Installations created before 3.2.0 keep using ~/.codex/ai-dlc-memory automatically. Schema migrations are versioned and run in place with a backup first. Deeper dives: Architecture · Learning model · Agent evolution.

The 26 agents

Area	Agents
Coordination	orchestrator, planner-requirements, product, research, knowledge-curator
Architecture & build	principal-architect, backend, frontend, data-database, integration-api, migration, ai-llm, ux-product-design, developer-experience
Quality	code-review, qa-verification, performance
Security	security, threat-model, red-team, governance-compliance
Operations	platform-sre, observability, incident-response, release, cloud-finops

Contracts live in contracts/agent_contracts.json and guides in src/ai_dlc_mcp/agents/. In Claude Code they are also available as ai-dlc-* subagents.

Commands

Setup installs these commands into the persistent uv tool environment (uv tool dir):

Command	Purpose
ai-dlc-admin	human governance: candidates, rules approve/reject, learnings verify, releases, purge, merge-projects, export, backup, audit
ai-dlc-install install | uninstall | doctor	register, remove or check the MCP server in Codex, Claude Code and VS Code
ai-dlc init	add a per-repository config (--project-mcp also writes .mcp.json)
ai-dlc doctor	check the local installation

Advanced: swarms, Mission Control, hooks and SDKs

None of these are needed for the memory workflow above.

• Swarms (ai-dlc swarm …): persistent, dependency-aware task graphs with atomic leases and isolated git worktrees.
• Mission Control (ai-dlc mission-control): local visual control plane, loopback-only by default.
• Hooks: session, prompt, post-tool and stop events through the bundled plugin; they fail open.
• SDK clients for Python and TypeScript talk to an MCP Streamable HTTP server only when you point them at one (AI_DLC_REMOTE_URL).
• Federation is off unless AI_DLC_ENABLE_FEDERATION=1.

See Advanced capabilities. The former agentshield command still works as a deprecated alias of ai-dlc.

Security model

• Write path: secrets (API keys, tokens, private keys, connection strings) are rejected before storage.
• Read path: prompt-injection-like content is flagged and withheld from context packs; context is delivered as data, not instructions.
• Human-only gates: approvals and purges need an interactive terminal and a typed confirmation, which agents running non-interactive shells cannot satisfy.
• Network: none by default. Mission Control binds to loopback and rejects non-loopback hosts.

Report vulnerabilities privately; see SECURITY.md.

FAQ

Does my code or memory leave my machine?

The MCP server stores memory locally in SQLite and has no telemetry. Your AI client may send retrieved context to its model provider under that client's configuration. Federation and remote SDK connections are opt-in.

Do Claude Code and Codex share the same memory?

Yes. Hosts configured with the default data directory use the same database, so a decision recorded in one is available in the others.

I installed an earlier version. Where is my memory?

Still in ~/.codex/ai-dlc-memory, and it keeps being used; nothing is moved. To adopt the new location, close your editors and move that folder to ~/.ai-dlc/memory.

Will it overwrite my existing agent or MCP configuration?

The installer edits its own marked block or server entry in AGENTS.md, CLAUDE.md and MCP configuration, and keeps timestamped backups.

Can an agent approve its own rules?

No. Agents can propose rules; approval requires ai-dlc-admin in an interactive terminal. Specialist contracts guide the host agent; they do not train or replace its model.

Documentation

Guide	Contents
Installation	Setup, updates, backups and uninstall
Architecture	MCP hosts, SQLite, retrieval and trust boundaries
Learning model	Evidence, verification and knowledge lifecycle
Agent evolution	Evaluation gates, releases and drift
Advanced capabilities	Swarms, Mission Control, hooks and opt-in features
Benchmarks	Reproducible measurement protocol
Security policy	Security controls and vulnerability reporting
Changelog	Release history

Uninstall

Run the installed ai-dlc-install uninstall --no-cli command before uv tool uninstall ai-dlc-engineering-os. Memory is preserved. See uninstall details, including source-script installs.

Contributing

Issues and pull requests are welcome. See CONTRIBUTING.md (DCO sign-off: git commit -s) and the changelog. Not affiliated with Anthropic, OpenAI, Microsoft or GitHub.

---

Português

Seus agentes de código esquecem tudo entre uma sessão e outra. O AI-DLC Engineering OS resolve isso.

Memória de engenharia local e governada + 26 agentes especialistas para Claude Code, Codex e VS Code, via MCP. Decisões, incidentes e padrões do projeto continuam disponíveis na próxima sessão e em qualquer editor configurado, e nada vira "regra" sem evidência e aprovação humana.

Gratuito e open source (MIT). O servidor de memória roda localmente, sem conta, telemetria ou serviço hospedado. Seu cliente de IA controla o envio do contexto recuperado ao provedor do modelo; conexões remotas são opcionais.

Instalação

Instale o uv. No Windows, macOS ou Linux, execute:

uvx ai-dlc-engineering-os setup

Escolha os editores, revise as alterações e confirme. O assistente instala um ambiente persistente, faz backups, registra o MCP e verifica os editores escolhidos. Reinicie o editor e pergunte: "O que já sabemos sobre este projeto antes de eu mexer no fluxo de pagamento?"

Pacote disponível no PyPI. Veja o guia de instalação para atualização e desinstalação.

Por que usar

• Memória que continua verdadeira: decisões, lições, incidentes e resultados, com busca textual e vetorial local. O repositório e os testes sempre prevalecem sobre a memória.
• Seguro por padrão: segredos são recusados, conteúdo com cara de prompt injection é isolado, e há exclusão definitiva (LGPD).
• Governança humana: agentes propõem; só humanos aprovam (ai-dlc-admin, em terminal interativo).
• Mesma memória no Claude Code, Codex e VS Code da sua máquina.
• Provado, não prometido: demo e benchmarks reproduzíveis com um comando (veja Proof).

Desinstalação: siga o guia. A memória é preservada.

Criado por José Camargo · LinkedIn · Instagram @jcamarguinho