Download the PHP package lindor03/project-memory without Composer

lindor03/project-memory

Local-first code intelligence for Laravel. The package indexes one application into that application's own database: modules, symbols, dependencies, architectural knowledge, change history, index runs, and retrieval metrics. Source code stays authoritative. The index is advisory. No LLM, embedding provider, or Laravel Boost installation is required.

Package version: 1.3.0.

Requirements

Do not treat the Composer constraint as a claim that every PHP 8.2+ or Laravel 12 patch release has been tested.

Install from a local path

From the consuming application:

Composer path repositories symlink or junction the package. Each application still gets its own index database.

Install from a private Git repository

Initialize and tag the package locally before any remote exists. This repository does not publish a tag or create a remote as part of installation.

In the consuming application, point Composer at the private repository and require the tag:

Use the real Git URL only after a remote exists. Do not commit credentials.

Initialize

project-memory:install does the following:

1. Checks PHP 8.2+ and Laravel 12.
2. Publishes config/project-memory.php only when that file is absent.
3. Opens the configured memory connection. For the SQLite default it creates the local file if needed.
4. Detects an existing schema.
5. On an empty memory database, applies the package's additive migrations. On a database that already has memory tables but is missing a later column, it prints guidance and waits for --migrate.
6. Leaves existing rows in place. It does not call migrate:fresh, migrate:rollback, or db:wipe.
7. Succeeds again when the schema is already current.

--migrate applies pending up migrations on the memory connection only. --cursor adds a portable MCP server and a Cursor rule only when those entries are absent. Existing MCP servers, including Laravel Boost, and an existing .cursor/rules/project-memory.mdc are left untouched.

Host php artisan migrate does not load these migrations, so they are not applied to the application database. ai:scan and ai:sync apply any still-pending additive migrations on the memory connection before indexing.

Database

The default connection name is project_memory. The default driver is SQLite and the default file is storage/app/project-memory.sqlite inside the consuming application. That file is not the application database. MemoryConnection refuses to use the application database, including two SQLite paths that resolve to the same file.

The package does not create or alter MySQL or MariaDB databases. To use a server, create an empty database yourself, then set:

In Docker, PROJECT_MEMORY_DB_HOST is the database service name reachable from the application container. One memory database belongs to one application checkout. A later scan refuses a stored repository path that does not match the current application. Use a different database for a different checkout.

blocked_databases is an optional extra deny list in config. It is empty in the package. Put application-specific database names there when a shared server has databases that must never receive this index.

Back up ai_knowledge and ai_change_sets before an upgrade. Migrations add the extraction cache and edge provenance columns. They do not delete knowledge rows or source hashes. ai:doctor is read-only.

Modules

config/project-memory.php accepts a modules map. Each entry has a key, name, and paths relative to the scan base (the application root unless scan.base_path is set). Paths may be directories or files. Unmapped files are grouped from their relative path. The package does not ship a starter-kit module map or knowledge catalog.

knowledge.catalog may suggest drafts. Synchronization never approves them.

Indexing and retrieval

ai:scan is the first full index. Later edits use ai:sync. Parse and publication failures keep the previous graph and record a failed run. ai:learn stores a draft unless --approve is present. Approval requires fresh indexed evidence. --revise=<knowledge_key> archives the previous version instead of overwriting it.

ai:evaluate measures retrieval against evaluation.scenarios in the published configuration. The package ships an empty list. ai:route and the context_route MCP tool choose evidence sources. They do not call Boost and they do not copy documentation into the memory database.

The benchmark script boots whichever Laravel application contains the current working directory, or PROJECT_MEMORY_APP_ROOT. It uses an in-memory SQLite database and generated files under the system temp directory:

MCP and Cursor

php artisan ai:mcp speaks newline-delimited JSON-RPC on stdin and stdout, and accepts Content-Length framing. Only protocol messages go to stdout. Writes require project-memory.mcp.allow_write=true and a literal boolean confirm. Shell execution is not exposed.

Portable project configuration, with php resolved from PATH and the working directory set to the application root:

php artisan project-memory:install --cursor writes that entry when project-memory is missing. It does not replace other servers.

When php is not on PATH, set command in the project file to the interpreter that should run this application. Do not copy that absolute path into the package. The launcher vendor/bin/project-memory-mcp uses the PHP binary that started it and finds artisan from the working directory or PROJECT_MEMORY_APP_ROOT:

On Windows the Composer bin proxy is vendor/bin/project-memory-mcp.bat. The server boots the application it finds. It does not read another application's files or database.

The same relative command works in Docker when the MCP client executes inside the application container and php is on that container's PATH. This release was not executed on Linux, macOS, or Docker.

Laravel Boost

Boost is suggested, not required. Project Memory boots, indexes, and serves MCP when Boost is absent.

When Boost is installed:

• Use Boost search-docs for official Laravel, Fortify, and Sanctum documentation.
• Use Project Memory for this application's symbols, modules, impact, and approved knowledge.
• context_route reports boost_installed and names search-docs only as the documentation tool. It does not duplicate a documentation index.
• MCP tool names are project_overview, module_context, symbol_lookup, impact_analysis, change_history, architecture_rules, index_status, context_route, sync_index, and record_knowledge. They do not replace Boost tools.
• --cursor does not remove an existing laravel-boost MCP server.

Upgrade

--migrate runs pending additive migrations on the memory connection. Knowledge rows, revisions, and source_hashes stay stored. A changed analyzer fingerprint can mark evidence stale on the next ai:sync; that does not delete the assertion. Review stale knowledge and revise it with current evidence. Do not reset the memory database to repair it.

Rollback

1. Restore the memory-database backup taken before the upgrade, especially ai_knowledge and ai_change_sets.
2. Require the previously verified package tag, for example composer require lindor03/project-memory:1.3.0 once that tag exists.
3. Prefer the backup over php artisan migrate:rollback. The edge-provenance migration refuses to roll back when duplicate declaration relationships exist, because dropping those columns would discard evidence. The down method also drops the extraction cache.

There is no supported automatic downgrade that rewrites knowledge into an older shape.

Release process

Semantic versions follow the Composer version field and ProjectMemory\PackageInfo::VERSION.

• Patch: diagnostics or indexing fixes that do not change stored schema or command contracts.
• Minor: additive schema, commands, or MCP fields. Version 1.3.0 is the portable package release.
• Major: removed commands, renamed MCP tools, or a migration that cannot preserve knowledge.

Before a tag:

1. Run composer test in this package.
2. Run the consuming application's focused project-memory tests and php artisan ai:doctor --json.
3. Confirm README compatibility matches versions that were actually executed.

Tag locally when the maintainer is ready. Do not push a remote or publish Packagist until that release is authorized:

Troubleshooting

Behavior preserved from the indexing engine

• Scans stay inside the configured base path, skip symlinks, and do not treat an excluded directory as a deletion.
• AST, route, Blade, and migration analyzers produce nodes and edges. Extraction cache stores summaries and metadata, not source bodies.
• Publication of the graph and knowledge invalidation commit together. A failed run keeps the previous graph.
• Impact analysis is a bounded reverse traversal plus containment.
• Context fits an estimated token budget. Token figures are UTF-8 byte estimates, not provider billing.
• MCP and CLI JSON keep the existing tool and field names. metrics remains additive.

Package tests

The suite uses Orchestra Testbench and does not open the development application's MySQL database. A fresh Laravel 12 application is an additional manual check described in the release process; it is not committed inside this package.

Compatibility matrix