Work Memory

Work Memory is a local-first development memory layer for VS Code and MCP clients. It stores cross-repository work history in SQLite, keeps an append-only JSONL archive for auditability, and exposes that shared memory through a local MCP server.

It is designed for one-machine-first use, but the codebase is portable and the memory store can be backed up, restored, or pointed at a synced location.

What It Does

• stores turns, notes, commands, files, entities, topics, and commit references
• imports local VS Code chat history from workspaceStorage
• serves that memory back into VS Code through MCP
• keeps retrieval deterministic with metadata-first filters before text search
• includes an RLM-inspired recursive recall mode with trajectory metadata and optional JSONL logs

Quick Start

python -m venv .venv
.venv\Scripts\activate
pip install -e .[dev]
work-memory init
work-memory-install-vscode

After that, reload VS Code and trust the work-memory MCP server when prompted.

For a published install from PyPI, use the package name work-memory-mcp:

pip install work-memory-mcp

VS Code Usage

If you use this from another VS Code window on the same machine, that other window can query the same history as long as it is configured to launch this MCP server. The memory data is already shared machine-wide by default under %LOCALAPPDATA%/work-memory, so the main requirement is MCP client configuration, not a separate database per window.

You do not need a global install just to use it across multiple VS Code windows on this computer. A workspace or user-level MCP config can point directly at this project's virtual environment:

{
  "command": "C:/Project/work-memory/.venv/Scripts/python.exe",
  "args": ["-m", "work_memory.mcp_server", "--transport", "stdio"]
}

If you want a more stable command that does not depend on this repo path, install the package into a dedicated environment and point VS Code at work-memory-mcp instead. That is a convenience and portability improvement, not a functional requirement.

To install it for all VS Code windows in your user profile, run:

work-memory-install-vscode

That script:

• creates a dedicated runtime under %LOCALAPPDATA%/work-memory/runtime
• installs or upgrades this project into that runtime
• updates your user-level VS Code mcp.json
• enables chat.mcp.autoStart

This repository also includes a workspace-scoped example at .vscode/mcp.json.

If you use VS Code Settings Sync, enable MCP server synchronization and the user-level mcp.json entry will follow you across machines. You still need the runtime installed on each machine, but the server registration itself can sync.

Architecture

The MVP keeps the boundary small:

• storage: initializes SQLite, writes immutable events, and appends raw JSONL records.
• extractor: derives tags, entities, and topics from text and metadata.
• retrieval: applies metadata filters first, then runs full-text search or topic matching.
• service: exposes a clean application interface that a future MCP wrapper can call.
• cli: provides local workflows without any cloud or LLM dependency.

Core concepts:

• project: repo-level identity (repo_name, repo_path)
• session: scoped work session (session_id, branch, source)
• event: immutable stored record (turn, note, command)
• artifact: file and commit references extracted from an event
• tags, entities, topics: normalized metadata for filtering and summarization

Setup

python -m venv .venv
.venv\Scripts\activate
pip install -e .[dev]

If you are installing from PyPI instead of from source:

pip install work-memory-mcp

Initialize storage. By default this now uses a machine-level directory under %LOCALAPPDATA%/work-memory so multiple repositories can share one memory store. Use --root only when you want an isolated project-local store.

work-memory init

Override the storage root for testing or local-only use:

work-memory --root C:/Project/work-memory init

Example Commands

Store a user turn:

work-memory store-turn \
  --repo-name repo-a \
  --repo-path C:/code/repo-a \
  --branch main \
  --session-id repo-a-2026-03-06 \
  --role user \
  --source cli \
  --text "How should we model auth and deployment memory?" \
  --tag architecture \
  --file src/auth.py

Store an assistant response:

work-memory store-turn \
  --repo-name repo-a \
  --repo-path C:/code/repo-a \
  --branch main \
  --session-id repo-a-2026-03-06 \
  --role assistant \
  --source cli \
  --text "Use metadata-first retrieval and keep raw history append-only." \
  --tag memory \
  --tag architecture

Store a command:

work-memory store-command \
  --repo-name repo-a \
  --repo-path C:/code/repo-a \
  --branch main \
  --session-id repo-a-2026-03-06 \
  --source terminal \
  --command "pytest -q" \
  --cwd C:/code/repo-a \
  --exit-code 0 \
  --file tests/test_auth.py

Store a note:

work-memory store-note \
  --repo-name repo-a \
  --repo-path C:/code/repo-a \
  --branch main \
  --session-id repo-a-2026-03-06 \
  --source manual \
  --title "handoff" \
  --text "Auth migration is blocked on Discord thread decisions." \
  --tag auth \
  --tag handoff

Project-scoped recall:

work-memory recall-project \
  --repo-name repo-a \
  --query auth \
  --tag architecture

RLM-inspired recursive recall:

work-memory recursive-recall \
  --query discord \
  --max-depth 2 \
  --branch-limit 3 \
  --log-dir ./logs

Cross-project recall with explicit filters:

work-memory recall-cross-project \
  --query discord \
  --topic auth \
  --entity Discord \
  --repo-name repo-a \
  --repo-name repo-b

Summarize a topic:

work-memory summarize-topic --topic memory --limit 10

Discover likely VS Code Copilot Chat sources on this machine:

work-memory discover-sources --limit 20

Import structured VS Code chat sessions as reconstructed user/assistant turns:

work-memory import-vscode-sessions --limit 200

Import discovered VS Code Copilot Chat resources into the shared store:

work-memory import-vscode-copilot --limit 200

Import an exported JSONL, JSON, or text transcript:

work-memory import-path \
  --path C:/exports/copilot-chat.jsonl \
  --repo-name imported-chat \
  --repo-path C:/imports/copilot-chat \
  --session-id import-2026-03-06 \
  --source exported-chat \
  --format auto

Import a VS Code chat session export or copied chatSessions/*.jsonl file with explicit structured reconstruction:

work-memory import-path \
  --path C:/exports/session.jsonl \
  --repo-name imported-chat \
  --repo-path C:/imports/session \
  --session-id import-2026-03-06 \
  --source exported-chat \
  --format vscode-chat-session

Seed Data

python scripts/seed_sample.py

That script creates example records across multiple repositories so you can validate project-scoped and cross-project retrieval.

Storage Layout

• %LOCALAPPDATA%/work-memory/work_memory.db: shared SQLite database by default
• %LOCALAPPDATA%/work-memory/raw_events.jsonl: append-only archive by default
• data/work_memory.db: used only when --root points at a project root
• data/raw_events.jsonl: used only when --root points at a project root

Raw history is never mutated after append. Search and recall operate from SQLite while the JSONL archive remains transparent for debugging or replay.

Current Import Reality

The importer can now do two different things for VS Code data:

• import-vscode-sessions reconstructs structured chat sessions from workspaceStorage/*/chatSessions/*.jsonl into user and assistant turns.
• import-vscode-copilot preserves Copilot chat resource files as raw imported notes for auditability and fallback coverage.

Repo inference now prefers workspace.json, which means single-folder VS Code workspaces can usually map imported session history back to the real repo path instead of a workspace hash.

MCP Server

Run the MCP server over stdio:

work-memory-mcp --transport stdio

Run it over streamable HTTP for shared local access:

work-memory-mcp --transport streamable-http --host 127.0.0.1 --port 8000 --json-response

The MCP layer exposes tools for:

• memory_status
• search_memory
• recall_project_memory
• recall_cross_project_memory
• recursive_recall_memory
• summarize_memory_topic
• store_memory_note
• discover_memory_sources
• import_vscode_sessions
• import_vscode_copilot_resources

It also exposes a memory://status resource and a memory_query_plan prompt.

For a local MCP client that supports stdio, point it at:

{
  "command": "C:/Project/work-memory/.venv/Scripts/python.exe",
  "args": ["-m", "work_memory.mcp_server", "--transport", "stdio"]
}

GitHub And Other Machines

The codebase is portable. It is not bespoke to only this machine, but the stored memory is local-first, which means the database and JSONL archive live on each machine unless you explicitly copy or sync them.

What GitHub gives you:

• backup and version history for the code
• a simple install source for other machines
• a shareable project other people can clone and run

What GitHub does not give you automatically:

• sync of %LOCALAPPDATA%/work-memory/work_memory.db
• sync of %LOCALAPPDATA%/work-memory/raw_events.jsonl

To use this on another computer:

git clone <your-repo-url>
cd work-memory
python -m venv .venv
.venv\Scripts\activate
pip install -e .[dev]
work-memory init

If you want to preserve existing history on a second machine, copy the storage directory or set WORK_MEMORY_HOME to a synced location before starting the server.

Examples:

$env:WORK_MEMORY_HOME = "D:/synced/work-memory"
work-memory init
work-memory-mcp --transport stdio

Portable backup and restore helpers are included:

work-memory-backup --destination C:/backups
work-memory-restore --backup C:/backups/work-memory-backup-20260306-120000.zip

If you want true cross-machine continuity instead of occasional backups, set WORK_MEMORY_HOME to a synced folder on each machine before starting the MCP server.

$env:WORK_MEMORY_HOME = "D:/synced/work-memory"
work-memory init
work-memory-install-vscode

PyPI Publishing

This repository is configured for GitHub Actions Trusted Publishing to PyPI.

Workflow file:

• .github/workflows/publish-pypi.yml

Trusted Publisher values for PyPI:

• owner: joe02740
• repository: work-memory
• workflow name: publish-pypi.yml
• environment name: leave blank

Release flow:

1. Add the Trusted Publisher in the PyPI project settings for work-memory-mcp.
2. Push this repository state to GitHub.
3. Create a GitHub release such as v0.1.0.
4. GitHub Actions builds dist/* and publishes the package to PyPI.

Recommended Distribution Path

The practical path is now:

1. Push the code to GitHub.
2. Keep the local memory store out of git.
3. Publish work-memory-mcp to PyPI from a GitHub release.
4. Install from PyPI or from source on each machine.
5. If you later want cross-machine history sync, add explicit export/import or a synced storage location.

That keeps the project local-first while still making it reusable by you and other people.