A persistent project knowledge store that gives AI assistants memory across sessions. Delivered as an MCP (Model Context Protocol) server.

ProjectMemory maintains a SQLite database of project-specific knowledge — conventions, decisions, known issues, lessons learned. AI assistants read from it at session start and write to it when they learn something new. Over time, the database builds a growing understanding of your project.

Download the latest release for your platform from Releases:

• ProjectMemory-win-x64.zip — Windows
• ProjectMemory-linux-x64.tar.gz — Linux
• ProjectMemory-osx-arm64.tar.gz — macOS (Apple Silicon)

Extract and configure in ~/.copilot/mcp-config.json (user-level) or .github/mcp.json (repo-level):

{
  "mcpServers": {
    "project-memory": {
      "command": "/path/to/ProjectMemory.Server",
      "args": []
    }
  }
}

Publish a single-file executable:

git clone https://github.com/micsh/ProjectMemory.git
cd ProjectMemory
dotnet publish src/ProjectMemory.Server -c Release -r win-x64 -o publish/win-x64
# Also available: linux-x64, osx-arm64

Configure in ~/.copilot/mcp-config.json (user-level) or .github/mcp.json (repo-level):

{
  "mcpServers": {
    "project-memory": {
      "command": "C:/path/to/ProjectMemory.Server.exe",
      "args": []
    }
  }
}

Pack and install as a global dotnet tool:

dotnet pack src/ProjectMemory.Server -c Release -o nupkg
dotnet tool install -g --add-source ./nupkg ProjectMemory

Then configure:

{
  "mcpServers": {
    "project-memory": {
      "command": "project-memory",
      "args": []
    }
  }
}

Update later with:

dotnet tool update -g --add-source ./nupkg ProjectMemory

{
  "mcpServers": {
    "project-memory": {
      "command": "dotnet",
      "args": ["run", "--no-build", "-c", "Release", "--project", "/path/to/ProjectMemory/src/ProjectMemory.Server"]
    }
  }
}

Build Release first: dotnet build -c Release

The database is created at .project-memory/memory.db in the working directory. Override with PROJECT_MEMORY_DB environment variable.

For best results, add this to your copilot instructions:

You have access to a project memory database via the project-memory MCP server.
- At the start of work, call get_context() to load project-specific knowledge.
- When the user corrects you, call record_lesson() with what you learned.
- When you discover a project-specific pattern, call project_store() to remember it.
- When you find stored knowledge is wrong, call project_forget() to remove it.

Knowledge and lessons can be scoped to specific files using glob patterns:

• * — project-wide (default)
• src/Auth/* — everything in Auth directory
• tests/**/*Integration* — integration test files
• *.fsproj — all F# project files

When get_context(scope: "src/Auth/Login.fs") is called, it returns up to 10 items ranked by:

1. Scope precision — items scoped to matching patterns rank above project-wide (*) items
2. Confidence — higher confidence ranks first
3. Recency — more recently updated items break ties

1. Lesson capture — The assistant records lessons when corrected. Fuzzy deduplication (Jaccard similarity >70%) prevents near-duplicate entries. When the active lesson count exceeds 30, consolidation runs automatically every 5th addition.
2. Consolidation — Merges similar lessons (>80% Jaccard via FTS5 pre-filtering), promotes lessons with high recurrence (≥5) and confidence (≥0.7) to knowledge, prunes stale entries (>30 days, single occurrence).
3. Confidence tracking — Each time get_context surfaces an item, its confidence increases by 0.05 (capped per session). Items not surfaced in the last 5 sessions decay by 0.03 per consolidation. Items that drop below 0.2 confidence are auto-archived.
4. Graduation — Knowledge entries reaching confidence ≥0.75 and surfaced in ≥5 distinct sessions are automatically appended to .github/copilot-instructions.md under an auto-managed section. Configurable via PROJECT_MEMORY_INSTRUCTIONS_FILE env var.
5. Un-graduation — If a graduated item's confidence later decays below the archive threshold, its instruction is automatically removed from the instructions file.

dotnet build
dotnet test
dotnet build -c Release          # For MCP / distribution

ProjectMemory.Server/
├── Schema.fs           # DDL, migrations (v1→v5), schema versioning
├── Similarity.fs       # Jaccard text similarity (pure function)
├── Formatting.fs       # Context output formatting with token budgeting (pure function)
├── InstructionsFile.fs # Graduation file I/O (pure function)
├── Database.fs         # ProjectMemoryDb class — persistence layer (read/write, no domain logic)
├── DomainService.fs    # Domain orchestration — Consolidate, Graduate, Ungraduate, auto-triggers
├── Tools.fs            # MCP tool adapter layer with input validation
├── Resources.fs        # MCP resource endpoints (memory://context)
└── Program.fs          # Host setup, DI, STDIO transport

The server uses STDIO transport — the MCP client (Copilot CLI, VS Code) launches it as a subprocess and communicates via stdin/stdout. All logging goes to stderr.