SQLite Project Memory MCP
SQLite 项目记忆 MCP
基于 SQLite 的 MCP 服务器,用于将项目记忆存储为对图友好的关系型核心。
为什么选择本项目
将权威状态保存在 SQLite 中(单一事实来源)。
提供图语义(
entities、relationships、content、tags)。支持结构化读取(
query_view)和显式生成的 Markdown 视图。专为 AI 代理安全的工作流而设计。
Related MCP server: @contextable/mcp
核心模型
entities(实体)attributes(属性)relationships(关系)content(内容)tags(标签)events、snapshots、snapshot_entities(审计/历史记录)
支持的 sync-document 目标
核心文档:
architecture、decisions、plan、notes、roadmap扩展文档:
kpi、okr、strategy、risk、issue、epic、story、feature、milestone、release、dependency、objective、initiative、metric、capability、assumption、problem_statement、retrospective、action_item
安装
将仓库克隆到项目的 sqlite-mcp 子文件夹中,然后运行一次安装程序。
脚本会自动检测其位置,将所有运行时文件放置在 Project Memory/ 中,并写入 .vscode/mcp.json。
无需安装 Python。 安装程序会下载 uv 并使用它自动获取固定的 CPython 3.12.9 解释器。解释器缓存于 Project Memory/.uv/python/,虚拟环境创建在 Project Memory/.venv 中。
如果 uv 下载失败(例如在防火墙后),安装程序将回退到机器上已有的任何 Python 3.11+ 环境。
重要提示:请在项目根目录(父文件夹)打开 VS Code,而不是在 sqlite-mcp 子文件夹中打开。
MCP 配置会写入项目根目录的 .vscode/mcp.json。
Windows (PowerShell)
# from your project root
git clone https://github.com/WebRTCGame/SQLITE-MCP.git sqlite-mcp
.\sqlite-mcp\install.ps1可选标志:
-LogFile install.log— 保存完整记录以供调试-AppendInstructions— 将代码片段附加到建议的指令文件中(幂等)
Linux / macOS
# from your project root
git clone https://github.com/WebRTCGame/SQLITE-MCP.git sqlite-mcp
chmod +x ./sqlite-mcp/install.sh
./sqlite-mcp/install.sh可选标志:
--log-file install.log— 保存完整记录以供调试--append-instructions— 将代码片段附加到建议的指令文件中(幂等)
更新方法: 嵌套安装成功后,脚本位于 Project Memory/ 下,因为 sqlite-mcp 检出目录已被移动到此处。请从项目根目录使用 Project Memory\install.ps1 (Windows) 或 Project Memory/install.sh (Linux/macOS) 重新运行安装程序。对于就地/开发者安装,请重新运行相同的命令。
仅使用 pip(开发者/高级)
用于您希望自行管理环境的情况(需要 Python 3.11+):
python -m venv .venv
source .venv/bin/activate # or .venv\Scripts\Activate.ps1 on Windows
pip install -e .或者直接使用 uv(无需系统 Python — uv 会自动下载 Python):
uv venv --python 3.12.9 .venv
uv pip install -e .完整示例:
./sqlite-mcp/install.sh启动服务器
& ".\Project Memory\.venv\Scripts\python.exe" -m sqlite_mcp_server路径
Project Memory/.venv— 虚拟环境(固定为 Python 3.12.9)Project Memory/.uv/bin/uv[.exe]— uv 运行时管理器(由安装程序下载)Project Memory/.uv/python/— 固定 CPython 解释器(由 uv 在首次运行时获取)Project Memory/pm_data/project_memory.dbProject Memory/pm_exports.vscode/mcp.json.vscode/settings.json可能也会由 Copilot/VS Code 自动创建,其中包含sqlite-project-memory的chat.mcp.serverSampling条目;这是预期行为。
要覆盖固定的 Python 版本:
$env:SQLITE_MCP_PYTHON_VERSION = "3.13.0"; .\sqlite-mcp\install.ps1SQLITE_MCP_PYTHON_VERSION=3.13.0 ./sqlite-mcp/install.shCLI 工具
sqlite-project-memory-admin bootstrap-self --repo-root .
sqlite-project-memory-admin project-state
sqlite-project-memory-admin health
sqlite-project-memory-admin sync-document architecture --input-path architecture.md
sqlite-project-memory-admin sync-document decisions --input-path decisions.md
sqlite-project-memory-admin sync-document roadmap --input-path roadmap.md
sqlite-project-memory-admin export-views --user-requested --request-reason "User asked for a roadmap export" --require-existing-dir exports todo roadmap architecture
sqlite-project-memory-admin export-views --user-requested --request-reason "User asked for refreshed generated docs" --force todo roadmap architecture
sqlite-project-memory-admin export-json --output-path exports/project_memory.snapshot.json
sqlite-project-memory-admin import-json --input-path exports/project_memory.snapshot.json配置
环境变量:
SQLITE_MCP_PROJECT_ROOTSQLITE_MCP_DB_PATHSQLITE_MCP_EXPORT_DIRSQLITE_MCP_TRANSPORT(stdio或streamable-http)SQLITE_MCP_LOG_LEVEL(默认为INFO)SQLITE_MCP_LOG_FORMAT(json或text)
安装后钩子
创建
.scripts/post_install.ps1(PowerShell) 或.scripts/post_install.sh。如果存在,安装脚本会执行此钩子。
快速入门 (Windows)
将仓库克隆到您的项目中。
运行安装程序:
git clone https://github.com/WebRTCGame/SQLITE-MCP.git sqlite-mcp .\sqlite-mcp\install.ps1安装后,确保在父项目根目录(例如
C:\CODE\TestProject)打开 VS Code,而不是C:\CODE\TestProject\sqlite-mcp。激活运行时虚拟环境:
& ".\Project Memory\.venv\Scripts\Activate.ps1"运行自检:
sqlite-project-memory-admin --db-path "Project Memory/pm_data/project_memory.db" project-state sqlite-project-memory-admin --db-path "Project Memory/pm_data/project_memory.db" health启动服务器:
python -m sqlite_mcp_server
快速入门 (Linux/macOS)
将仓库克隆到您的项目中。
运行安装程序:
git clone https://github.com/WebRTCGame/SQLITE-MCP.git sqlite-mcp chmod +x ./sqlite-mcp/install.sh ./sqlite-mcp/install.sh安装后,确保在父项目根目录打开 VS Code,而不是
sqlite-mcp子文件夹。激活运行时虚拟环境:
source "Project Memory/.venv/bin/activate"运行自检:
sqlite-project-memory-admin --db-path "Project Memory/pm_data/project_memory.db" project-state sqlite-project-memory-admin --db-path "Project Memory/pm_data/project_memory.db" health启动服务器:
python -m sqlite_mcp_server
注意事项
render_markdown_views/export_markdown_views需要user_requested=true和request_reason。sync-document是锚定的(路线图状态中没有非结构化的自由文本记录)。project_state和health是主要的诊断指标。
AI 优先设计
通过 MCP 工具编写实体/内容。
仅在用户明确请求时生成文档。
将 AI 指令添加到您的项目
运行安装程序后,将下面的代码片段粘贴到您的 AI 指令文件中(copilot-instructions.md、AGENTS.md、CLAUDE.md 或等效文件)。安装程序也会将此片段打印到终端以作提醒。
## SQLite Project Memory
This project uses the **`sqlite-project-memory` MCP server** as the single source of truth for all project state.
### Mandatory rules
- **NEVER** create, modify, or delete `.md` files for notes, tasks, decisions, roadmaps, or architecture.
- **ALWAYS** use `sqlite-project-memory` MCP tools to store and retrieve project knowledge.
- `README.md` is the **only** markdown file you may edit directly.
- Files in `Project Memory/pm_exports/` are read-only derived artifacts — never edit them directly.
### Tool usage guide
| Intent | Tool to use |
|--------|-------------|
| Add a task / todo | `create_entity` (entity_type: `task`) |
| Record a decision | `create_entity` (entity_type: `decision`) |
| Document a component | `create_entity` (entity_type: `component`) + `write_content` |
| Add a note to an item | `write_content` or `append_content` |
| Query project state | `query_view`, `list_entities`, `get_entity` |
| Search knowledge | `search_content` |
| Link two items | `add_relationship` or `connect_entities` |
| Export to markdown | `export_markdown_views` with `user_requested: true` — only when explicitly asked |
### First action each session
Before making changes, call `get_project_context` to confirm the database path and project root, then call `get_recent_activity` or `query_view` to orient yourself.使用门控检查清单
为了获得最佳可靠性,请确认以下所有事项:
.vscode/mcp.json包含sqlite-project-memory服务器。已为此工作区接受了 VS Code MCP 批准/信任提示。
聊天处于代理模式(非普通聊天/编辑模式)。
Project Memory代理在代理下拉菜单中可用。/sqlite-project-memory技能在斜杠命令中可用。您的项目指令文件包含 SQLite 项目记忆代码片段。
每次会话开始时使用
get_project_context,然后使用get_recent_activity或query_view。
安装程序会打印一份“使用门控报告”,其中包含它可以自动验证的门控的 PASS 或 ACTION REQUIRED 状态。
如果工具仍然没有出现,请先完全重新加载或重启 VS Code 并开始新的代理聊天会话。
如果重启后服务器没有运行,请从命令面板运行 MCP: Start Server,选择 sqlite-project-memory,然后重启聊天会话。
Available Tools
41 toolsadd_relationshipC
Create a typed graph edge between two entities.
Inputs: relationship_id (required), from_entity (required), to_entity (required), relationship_type (required). Returns: object
append_contentD
Add a piece of content to an entity.
Inputs: entity_id (required), content_type (required), body (required), content_id. Returns: object
apply_performance_tuningC
Tune SQLite settings for throughput during heavy project-memory workloads.
Inputs: journal_mode, synchronous, temp_store, cache_size, +2 more. Returns: object
archive_entityA
Archive an entity without deleting its history or related project memory.
Inputs: entity_id (required), reason, archived_status. Returns: object
bootstrap_project_memoryC
Initialize a project root entity and standard memory-area anchor entities.
Inputs: project_id (required), project_name (required), description, tags. Returns: object
connect_entitiesC
Create a relationship if missing, otherwise return the existing edge.
Inputs: from_entity (required), to_entity (required), relationship_type (required), relationship_id. Returns: object
create_entityC
Create a new authoritative project entity.
Inputs: entity_id (required), entity_type (required), name, description, +3 more. Returns: object
create_snapshotC
Create a named snapshot of current project state.
Inputs: snapshot_id (required), name (required), description, entity_ids. Returns: object
delete_entityA
Delete an entity with guardrails; non-forced deletion requires prior archiving and no critical dependents.
Inputs: entity_id (required), force. Returns: object
delete_relationshipB
Delete a relationship by id and record the removal in project history.
Inputs: relationship_id (required). Returns: object
export_markdown_viewsB
Write generated markdown views only after an explicit user request for a human-readable artifact.
Inputs: view_names, output_dir, overwrite, require_existing_dir, +2 more. Returns: object
find_similar_entitiesB
Find likely duplicate or related entities before creating a new one.
Inputs: name (required), entity_type, limit. Returns: result
get_architecture_summaryB
Return a compact architecture-oriented node and relationship summary.
Inputs: node_limit, relationship_limit, compact. Returns: object
get_database_healthB
Report likely duplicates, low-quality records, and retention pressure in project memory.
Inputs: limit. Returns: object
get_decision_logB
Return decisions and recent supporting note excerpts without requiring ad hoc SQL.
Inputs: limit, offset, compact. Returns: object
get_dependency_viewC
Return dependency-oriented graph data with stable compact fields.
Inputs: root_entity_id, max_depth, relationship_types, limit, +1 more. Returns: object
get_entityC
Fetch an entity and optionally include related metadata, content, and events.
Inputs: entity_id (required), include_related. Returns: object
get_entity_graphC
Traverse outward relationship dependencies from an entity.
Inputs: entity_id (required), max_depth, relationship_type, edge_limit, +2 more. Returns: object
get_or_create_entityB
Reuse an exact entity when it already exists or create one with a generated stable id.
Inputs: entity_type (required), name (required), entity_id, description, +3 more. Returns: object
get_project_contextA
Return the current project root, DB path, and export directory.
Inputs: no named parameters. Returns: object
get_recent_activityC
Return recent entities, content, and events to help an AI resume context quickly.
Inputs: limit, offset, compact. Returns: object
get_recent_reasoningC
Return recent reasoning excerpts for quick AI context recovery.
Inputs: limit, offset, compact. Returns: object
get_snapshotB
Fetch a named snapshot and its captured entities.
Inputs: snapshot_id (required). Returns: object
list_entitiesC
List entities with optional type, status, attribute, tag, and text filters.
Inputs: entity_type, status, attribute_key, attribute_value, +4 more. Returns: result
list_relationshipsC
List graph edges, optionally constrained to an entity and direction.
Inputs: entity_id, relationship_type, direction, limit. Returns: result
list_viewsB
List available view names for query_view discovery.
Inputs: no named parameters. Returns: result
merge_entitiesB
Merge a duplicate source entity into a target entity with deterministic conflict handling.
Inputs: source_entity_id (required), target_entity_id (required), attribute_conflict. Returns: object
prune_content_retentionC
Prune older reasoning/log content while keeping the most recent records per entity and type.
Inputs: content_types, keep_latest, entity_id, dry_run. Returns: object
query_viewB
Query a named SQL view for generic reporting and model projections.
Inputs: view_name (required), params. Returns: object
refresh_task_summaryB
Rebuild a task summary materialized table for faster open task queries.
Inputs: no named parameters. Returns: object
render_markdown_viewsA
Render markdown document views only after an explicit user request for a human-readable artifact.
Inputs: view_names, user_requested, request_reason. Returns: object
resolve_entity_by_nameC
Resolve a human-readable name to an existing entity when possible.
Inputs: name (required), entity_type, limit. Returns: object
run_read_queryC
Run a constrained read-only SQL query for diagnostics, analytics, and ad hoc retrieval.
Inputs: sql (required), parameters, limit. Returns: object
search_contentC
Search narrative content using FTS5 when available.
Inputs: query (required), limit. Returns: result
server_infoB
Return server metadata and schema settings.
Inputs: no named parameters. Returns: object
set_project_rootB
Switch project context (root + db + export path) and reconnect database.
Inputs: project_root (required). Returns: object
set_tagsC
Set or merge tag labels for an entity.
Inputs: entity_id (required), tags (required), replace. Returns: object
update_entityC
Update an existing entity's top-level state.
Inputs: entity_id (required), name, description, status. Returns: object
upsert_attributesB
Create or update flexible metadata for an entity.
Inputs: entity_id (required), attributes (required), replace. Returns: object
upsert_entityC
Create an entity if missing or merge updates into an existing entity.
Inputs: entity_id (required), entity_type (required), name, description, +5 more. Returns: object
write_contentC
Write narrative content in one unified API with append|replace semantics.
Inputs: entity_id (required), mode (required), content (required), content_type (required). Returns: object
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 41 tools
The tools cover distinct operations like entity management, graph relationships, content handling, and queries, but there is notable overlap in some areas. For example, create_entity, get_or_create_entity, and upsert_entity have overlapping purposes for entity creation, and append_content vs. write_content could cause confusion in content addition. Descriptions help clarify, but agents might misselect between these similar tools.
Most tools follow a consistent verb_noun or verb_adjective_noun pattern (e.g., create_entity, list_entities, get_entity_graph), with clear and descriptive names. There are minor deviations like server_info (noun_verb) and apply_performance_tuning (verb_noun_noun), but overall the naming is predictable and readable across the set.
With 41 tools, the count is excessive for a project memory server, leading to potential cognitive overload and inefficiency. While the domain is broad, many tools could be consolidated or streamlined (e.g., multiple entity creation tools, overlapping content tools), making it feel heavy and less well-scoped than ideal for agent use.
The tool set provides comprehensive coverage for project memory management, including full CRUD operations for entities and relationships, content handling, graph traversal, queries, snapshots, health checks, and context management. There are no obvious gaps; agents can perform all core workflows from initialization to archiving and reporting without dead ends.
Maintenance
Related MCP Connectors
Shared project memory for people and AI agents: documents, ERD, diagrams, boards, design mockups.
Project memory, documentation and controlled secret access for AI coding agents.
The system of record for AI agents — open-source (MIT) database where agents and people share data.
Project memory for coding agents: requirements, decisions, code graph and delivery telemetry.
Related MCP Servers
- AlicenseAqualityDmaintenanceA knowledge graph memory server using SQLite to provide persistent, isolated contexts for organizing information into searchable categories like work and personal projects. It features unique ID-based operations and a token-efficient serialization format designed to optimize interactions with LLMs.11MIT
- AlicenseNot gradedqualityDmaintenanceA persistent AI memory server that enables storage and retrieval of context and project artifacts across conversations. It features full-text search, version history, and automatic content chunking using local SQLite or hosted cloud storage.13 npmApache 2.0
- AlicenseNot gradedqualityDmaintenanceA persistent memory server for AI agents that stores structured notes in a local SQLite database with full-text search and graph-based relationships. It features 32 specialized tools for managing long-term context, including version history, automated TTL expiration, and complex filtering.12 npmMIT
- AlicenseNot gradedqualityDmaintenanceA database-backed MCP server that acts as a project memory bank, enabling AI assistants to store, retrieve, and search structured context like decisions, tasks, and architecture using SQLite and vector embeddings.Apache 2.0