Skip to content

setup.md

Latest commit

 

History

History
160 lines (109 loc) · 3.72 KB

setup.md

File metadata and controls

160 lines (109 loc) · 3.72 KB

Setup

Single machine

All components run locally. The daemon ingests conversations, digests them, and agents query the same local database.

1. Build and install

just install

This installs lore, lore-mcp, lore-tray, lore-server, and lore-explorer to ~/.local/bin/.

On macOS, you can also build the desktop app:

just bundle-macos
cp -r target/Lore.app ~/Applications/

2. Register the MCP server

claude mcp add --scope user memory -- lore-mcp

3. Start the daemon

Either launch the Lore desktop app (auto-starts the daemon), or:

lore daemonize

The daemon stages conversation turns every 30 seconds and runs consolidation every 2 hours. No further setup needed — it works out of the box using claude -p for extraction.

4. Verify

lore status       # should show "running"
lore roots        # list extracted knowledge (empty at first, populates after consolidation)
lore staged       # show conversations waiting to be digested

Central server

Multiple client machines push conversation data to a central server. Agents on any machine query the shared knowledge base.

Server setup

1. Build the server

cargo build --release -p lore-server -p lore-daemon
cp target/release/lore-server target/release/lore ~/.local/bin/

2. Start the server

lore-server --port 8080 --db /path/to/shared/memory.db

The server exposes:

  • /mcp — MCP tools over streamable HTTP (for agents)
  • /push — accepts staged turns from clients
  • /status — health check

3. Start the consolidation daemon on the server

The server needs a daemon to digest pushed turns into knowledge:

lore start --log-file /var/log/lore.log

This runs consolidation against all turns pushed by clients.

4. Register the remote MCP server on each client

claude mcp add --scope user memory -- npx mcp-remote http://server:8080/mcp

Or configure your MCP client to connect to http://server:8080/mcp directly if it supports streamable HTTP transport.

Client setup

1. Build and install the CLI

cargo build --release -p lore-daemon
cp target/release/lore ~/.local/bin/

2. Configure remote mode

Create ~/.lore/config.toml:

[remote]
url = "http://server:8080"
sync_interval_secs = 60

3. Start the client daemon

lore daemonize

The client daemon:

  • Stages new conversation turns every 30 seconds (instant, no API calls)
  • Syncs staged turns to the server every 60 seconds
  • Does not run consolidation or require Claude API access
  • Does a final sync on shutdown

4. Verify

lore status       # should show "running"
lore staged       # should be empty after sync (turns are on the server)

To manually trigger a sync:

lore sync http://server:8080

Docker deployment

The central server can also be deployed via Docker:

just docker-build
docker run -d \
  --name lore \
  -p 8080:8080 \
  -v lore-data:/data \
  -e ANTHROPIC_API_KEY=sk-... \
  lore-server

The container runs both lore-server (MCP over HTTP) and the consolidation daemon. Configuration is via environment variables:

Variable Default Description
ANTHROPIC_API_KEY — Required for consolidation. Server runs without it but won't digest.
LORE_CONSOLIDATION_INTERVAL 7200 Seconds between consolidation cycles
LORE_IDLE_THRESHOLD 300 Seconds before a session is eligible for digestion
LORE_EXTRACTION_MODEL claude-sonnet-4-20250514 Model for knowledge extraction
LORE_COMPRESSION_MODEL claude-haiku-4-5-20251001 Model for recursive summarization

Data is persisted at /data/memory.db inside the container.