Memory for your AI chats

The cheapest integration of all: Memory Layer already ships an MCP server, so giving your AI assistant durable memory across conversations is configuration, not code.

1. Create a chats project and seed it

memory remember --project chats --type user \
  --title "Working preferences" \
  --summary "Baseline preferences for AI assistants" \
  --note "Prefers concise answers with sources; working hours Europe/London; primary projects: memory-layer, rillforge."

Add anything an assistant should durably know: preferences, ongoing projects, decisions, people. memory ingest works too if you keep such notes as files.

2. Point Claude Desktop at it

// claude_desktop_config.json
{
  "mcpServers": {
    "memory-layer": {
      "command": "memory",
      "args": ["mcp", "run", "--project", "chats"]
    }
  }
}

Codex and other MCP clients take the same stdio command. The server is read-first by design: the assistant can query, resume, and inspect memories, but cannot write, delete, or reconfigure anything — your memory stays yours.

3. Use it

Ask your assistant things only your memory can answer:

"Check memory: what did I decide about the greenhouse irrigation?"

The MCP tools return cited memories, and the same honest-refusal contract applies — no memory, no invented answer.

Closing the loop

Since the MCP surface is read-only, write back at the end of a conversation yourself:

memory remember --project chats --type project \
  --title "Chat outcome: holiday planning" \
  --summary "Settled on the Dolomites for September" \
  --note "Week 2 of September; basecamp in Val Gardena; needs the cabin booked by July."

One project, every assistant: because the memory lives in Memory Layer rather than in any one vendor's chat history, Claude Desktop, Codex, and your CLI all read the same durable knowledge.