Continuum
Continuum is a local-first persistent memory and git/context server for AI agents — it stores memories in local SQLite, retrieves them, manages projects, and exposes read-only git info.
Memory CRUD: save (
memory_save), search (memory_search, lexical/FTS), fetch by id (memory_get), update (memory_update), delete with confirm (memory_delete), list recent (memory_recent)Project management: list (
project_list), register a local path (project_register), get by id (project_get), assemble context with git status + recent memories (project_context)Read-only git inspection:
git_status(with protected-branch detection),git_diff(staged/unstaged, truncated if large),git_log,branch_listMemory metadata: types (DECISION/FACT/etc.), scope, confidence, importance, summary, source, project association
Safety: destructive delete requires
confirm=true; git operations are read-only (no push/force-push exposed)No cloud/telemetry: all data stays in local
~/.continuumSQLite; embeddings via optional local Ollama
Provides read-only tools for interacting with local Git repositories, including checking status and protected branch, viewing diffs (truncated), listing recent commits, and listing local branches.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@Continuum remember that we decided to use SQLite for local memory"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Continuum
OS de engenharia com IA, local-first — memória persistente para agentes via MCP.
One memory. Every agent. Continuous engineering.
English summary · Privacidade · Licença MIT
O que é
O Continuum guarda decisões, contexto de projeto e estado útil para agentes de IA (Cursor e outros) num banco SQLite local. Expõe:
CLI
continuum(init, doctor, projetos, memória, git read-only)Servidor MCP stdio para o agente chamar ferramentas de memória e git
Tudo roda na sua máquina. Sem cloud Continuum, sem telemetria.
Related MCP server: AgentMemory Offline
Local-first e privacidade
Sem banco na nuvem — dados em
~/.continuum/(override:CONTINUUM_HOME)Zero telemetria — o Continuum não phone-home
Offline por padrão — nada sai da máquina sem ação explícita sua
Cursor / LLM — o modelo e a conta do Cursor (ou outro host) são separados do Continuum; o Continuum só oferece ferramentas MCP locais
Detalhes: PRIVACY.md.
~/.continuum/
config.toml
continuum.db
logs/
cache/
backups/Estado do projeto
Já funciona (MVP + roadmap A–G + polish):
SQLite local + CRUD de memória + busca lexical (FTS5)
Embeddings locais via Ollama + vector store SQLite (BLOB) + busca híbrida
Indexador de repositório (
continuum index/continuum project index)Context assembler (
project_contextcom seções compactas)Knowledge graph em SQLite (entities / relationships) + CLI/MCP + UI com filtros
Política de branch/commit (Conventional Commits, quality gates; sem force push)
GitHub PRs (token via
GITHUB_TOKEN/ keyring; merge só com--approve)Handoff de sessão + auto-detect DECISION/CONSTRAINT + supersede em conflito
Explorer web (
apps/webVue+Vite +continuum serveem:8787)App desktop Windows (
continuum desktop/ atalho com ícone)Projetos + CLI e MCP stdio + logging JSON local
Adiado: Neo4j, sync multi-dispositivo (ver abaixo), merge automático em main.
Publicação PyPI está preparada (uv build / uv publish) e exige token.
Requisitos
Python 3.12+
uv (desenvolvimento / clone)
Git no
PATH(ferramentas git)
Instalação
Usuário final (quando publicado no PyPI)
pip install continuum
# ou
uv tool install continuumDepois o comando continuum fica no PATH.
Desenvolvimento / clone (agora)
git clone https://github.com/matheusxdev/continuum.git
cd continuum
uv sync --extra devInstalação editável a partir do git (sem clonar manualmente):
uv pip install "git+https://github.com/matheusxdev/continuum.git"
# ou
pip install "git+https://github.com/matheusxdev/continuum.git"Embeddings com Ollama (recomendado)
Instale e inicie o Ollama.
Puxe um modelo de embedding:
ollama pull nomic-embed-textEm
~/.continuum/config.toml:
embedding_provider = "ollama"
embedding_model = "nomic-embed-text"
ollama_base_url = "http://127.0.0.1:11434"Confirme:
continuum doctor(Ollama + Embedding model OK).memory_savepassa a gerar embeddings;memory search --mode hybrid|semanticusa o vetor. Se o Ollama estiver offline, a busca degrada para lexical sem falhar.
Quick start
uv run continuum init
uv run continuum doctor
uv run continuum project add .
uv run continuum memory save "Decidimos continuar com Vue + Vite" --type DECISION
uv run continuum memory search "Vue"
uv run continuum index
uv run continuum git status
uv run continuum project contextCom o pacote já no PATH (pip install / uv tool install):
continuum init
continuum doctor
continuum project add .
continuum mcpCursor — configuração MCP
Recomendado (pacote instalado no PATH)
{
"mcpServers": {
"continuum": {
"command": "continuum",
"args": ["mcp"]
}
}
}Desenvolvimento com uv (caminho do clone)
Substitua o diretório pelo caminho absoluto do repositório na sua máquina:
Windows
{
"mcpServers": {
"continuum": {
"command": "uv",
"args": ["run", "--directory", "C:/path/to/continuum", "continuum", "mcp"]
}
}
}macOS / Linux
{
"mcpServers": {
"continuum": {
"command": "uv",
"args": ["run", "--directory", "/path/to/continuum", "continuum", "mcp"]
}
}
}Há um exemplo em .cursor/mcp.json no repositório (forma PATH).
Ferramentas MCP
Tool | Modo | Descrição |
| mutating | Persistir memória |
| read-only | Busca híbrida (lexical + semântica) |
| read-only | Só vetor (Ollama) |
| — | CRUD |
| mutating | Marcar memória como SUPERSEDED |
| — | Projetos + contexto em seções |
| — | Git seguro (sem force push) |
| read-only | Knowledge graph |
| mutating | GitHub PRs (merge exige |
| read-only | Resumo de handoff |
Explorer web (mínimo)
cd apps/web && npm install && npm run build
uv run continuum serve # http://127.0.0.1:8787
# ou só o front em dev:
cd apps/web && npm run dev # proxy /api → :8787App desktop (Windows)
Janela dedicada (WebView2) com a mesma UI de http://127.0.0.1:8787 — sem depender de uma aba do navegador, e pode ir para a barra de tarefas.
# 1) Build da UI (se ainda não tiver apps/web/dist)
cd apps/web; npm install; npm run build; cd ../..
# 2) Dependência pywebview + abrir
uv sync --extra desktop
uv run continuum desktop
# ou:
.\scripts\continuum-desktop.ps1O comando sobe continuum serve em segundo plano só se a porta 8787 estiver livre. Ao fechar a janela, encerra apenas o servidor que ele mesmo iniciou.
Fixar na barra de tarefas
# Gera assets/continuum.ico (se faltar) e cria Continuum.lnk na Área de trabalho
.\scripts\continuum-desktop.ps1 -InstallShortcut
# Também no Menu Iniciar:
.\scripts\continuum-desktop.ps1 -InstallShortcut -StartMenuAlsoDepois abra o atalho uma vez e, na barra de tarefas, botão direito → Fixar na barra de tarefas.
Ícone: assets/continuum.ico.
Requisito: WebView2 Runtime (já presente na maioria dos Windows 10/11).
Indexar repositório
Indexa README / markdown / código (respeita denylist: .env, node_modules, .git, secrets), grava chunks como memórias PROJECT_CONTEXT e gera embeddings quando Ollama está ativo. Incremental por hash.
uv run continuum index
uv run continuum project index
uv run continuum index --force # reindexar tudoSync multi-dispositivo (adiado)
Sync entre máquinas não entra no escopo atual: quebraria o modelo local-first (sem cloud Continuum, sem telemetria). Continuidade entre PCs fica para um desenho futuro explícito (export/import opt-in ou sync peer-to-peer), nunca automático. Use backup de ~/.continuum/ se precisar mover dados.
Layout do código
src/continuum/
apps/cli/ CLI (Typer)
apps/mcp/ MCP stdio
apps/desktop/ Shell pywebview (Windows)
apps/webapi/ API local do explorer
core/memory/ Memória + hybrid search
core/embeddings/ Ollama / null providers
core/vector/ SQLite BLOB vector store
core/graph/ Knowledge graph
core/context/ Context assembler
core/indexer/ Repository indexer
core/git/ Git + commit policy + quality gates
core/github/ GitHub PRs
core/handoff/ Handoff + auto-type + supersede
database/ Schema + migrations
config/ Paths, settings, logging
apps/web/ Vue + Vite explorer
assets/ continuum.ico
scripts/ Launchers (desktop.ps1 / .bat / generate_ico)
docs/ ADRsTestes e build
uv sync --extra dev
uv run pytest
uv buildArtefatos em dist/ (wheel + sdist).
Publicar no PyPI (quando tiver token)
Não publique sem UV_PUBLISH_TOKEN / credenciais configuradas.
uv build
uv publishAlternativa com Twine:
uv build
uv run twine upload dist/*Crie um token em https://pypi.org/manage/account/token/ e use TestPyPI primeiro se preferir:
uv publish --publish-url https://test.pypi.org/legacy/Spec / roadmap
CONTINUUM_SPEC.md — especificação original
CONTINUUM_COMPLETA.md — visão completa / roadmap
English summary
Continuum is a local-first AI Engineering OS: persistent agent memory over local SQLite (~/.continuum), with a Typer CLI and MCP stdio server for Cursor and other hosts. No Continuum cloud DB, no telemetry. The Cursor/LLM provider is separate from Continuum itself.
Install (when on PyPI): pip install continuum
Dev now: git clone → uv sync --extra dev → uv run continuum init
See PRIVACY.md for privacy guarantees. MIT licensed.
Available Tools
14 toolsbranch_listBRead-onlyIdempotent
List local git branches and mark protected ones.
Inputs: path. Returns: result
git_diffARead-onlyIdempotent
Git diff (unstaged by default; set staged=true for index). Large diffs are truncated.
Inputs: path, staged. Returns: result
git_logCRead-onlyIdempotent
Recent commit log.
Inputs: path, limit. Returns: result
git_statusBRead-onlyIdempotent
Git status with protected-branch detection (main/master/production).
Inputs: path. Returns: result
memory_deleteADestructiveIdempotent
Delete a memory. Requires confirm=true (high-risk).
Inputs: memory_id (required), confirm. Returns: result
memory_getCRead-onlyIdempotent
Get a memory by id.
Inputs: memory_id (required). Returns: result
memory_recentBRead-onlyIdempotent
List recent memories for the current or given project.
Inputs: limit, project_id. Returns: result
memory_saveC
Save a persistent memory (decision, fact, preference, etc.).
Inputs: content (required), scope, source, summary, +4 more. Returns: result
memory_searchBRead-onlyIdempotent
Search memories with lexical/FTS retrieval.
Inputs: query (required), limit, project_id. Returns: result
memory_updateB
Update fields of an existing memory.
Inputs: memory_id (required), content, summary, confidence, +2 more. Returns: result
project_contextCRead-onlyIdempotent
Return registered project context, git status, and recent memories.
Inputs: path, project_id, recent_limit. Returns: result
project_getBRead-onlyIdempotent
Get a project by id.
Inputs: project_id (required). Returns: result
project_listBRead-onlyIdempotent
List registered Continuum projects.
Inputs: no named parameters. Returns: result
project_registerCIdempotent
Register a local path as a Continuum project.
Inputs: name, path. Returns: result
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 14 tools
Each tool targets a distinct resource+action: the memory_* family cleanly separates save/search/get/update/delete/recent, project_get vs project_context differ by scope, and git_status/git_diff/git_log/branch_list cover orthogonal git queries. No two tools appear to do the same thing, so an agent can select confidently.
Most tools follow a consistent resource_verb snake_case pattern (memory_save, project_register, git_status). The lone deviation is branch_list, which drops the git_ prefix used by its sibling git tools, but overall readability remains high.
14 tools split evenly across three coherent domains (memory, project, git) is well-scoped for a context/memory server. Each tool earns its place without redundancy.
Memory CRUD is fully covered (save/get/update/delete/search/recent) plus project lifecycle and git inspection. Minor gaps: no project update/unregister/delete or branch checkout/creation, but core workflows are workable.
Related MCP Connectors
Persistent memory, knowledge graph, task management and more for AI agents via MCP.
Private-by-default, local-first memory/context/task orchestrator for MCP apps and agents.
Shared memory for your AI tools: tasks, notes and context every tool can read and write over MCP.
Cross-tool persistent memory and context for AI assistants over MCP.
Related MCP Servers
- AlicenseNot gradedqualityFmaintenanceLocal-first, auditable memory for AI agents. Provides durable context for MCP hosts with SQLite storage, CLI, and MCP tools for memory management.2Apache 2.0
- AlicenseNot gradedqualityDmaintenanceProvides local, project-aware memory for coding agents via MCP tools, lifecycle hooks, and a REST service, with no outbound network access.34,305 npmApache 2.0
- AlicenseNot gradedqualityCmaintenanceGives coding agents a local, SQLite-backed memory of a codebase, enabling them to query symbol impact, change history, and task scope through MCP while recording what actually changed after edits.MIT
- AlicenseBqualityDmaintenanceProvides local, explicitly scoped memory for coding agents via MCP, storing durable project knowledge in a per-repository SQLite database with tools to record, search, and retrieve context.5MIT