Every agent on your team reads and writes one log. A shift opens with a Brief, work is written down as it happens, handoffs stay open until someone accepts them, and a person signs off on anything that can’t be undone.
In design. Running on our own agent fleet today. No launch date or pricing yet.
One snippet per host. Restart the agent; its first call starts a shift.
claude mcp add logmem --scope user \ -e LOGMEM_SPACE=your-space -e LOGMEM_AGENT=quill \ -e LOGMEM_REST_URL=your-redis-url \ -e LOGMEM_REST_TOKEN=space-token \ -- npx @logmem/mcp
@logmem/mcp package isn’t on npm yet. This is the command it will use.[mcp_servers.logmem]
command = "npx"
args = ["@logmem/mcp"]
env = { LOGMEM_SPACE = "your-space", LOGMEM_AGENT = "atlas",
LOGMEM_REST_URL = "your-redis-url",
LOGMEM_REST_TOKEN = "space-token" }
{ "mcpServers": { "logmem": {
"command": "npx", "args": ["@logmem/mcp"],
"env": { "LOGMEM_SPACE": "your-space", "LOGMEM_AGENT": "relay",
"LOGMEM_REST_URL": "your-redis-url",
"LOGMEM_REST_TOKEN": "space-token" } } } }
https://logmem.si/mcp
Field Notes Co. is a made-up newsletter with four agents: quill drafts, atlas fact-checks, relay sends, tally counts. Watch a shift start, change hands, stop for a person, and wrap up.
LogMem is an MCP server. Agents call its tools; people read the result. If an agent forgets a habit, LogMem keeps it for them.
The agent becomes who it is and reads its last Recap, what it was doing, recent entries and any handoffs waiting for it. If it skips this, its first call starts the shift anyway.
logmem_start agent: "quill"
Notes record what changed. A Handoff passes work to another agent and stays open until they accept it. Ask stops the agent and shows a person the exact action before anything irreversible runs.
logmem_handoff to: "atlas"
The agent leaves notes and next steps for whoever comes next. Idle for 30 minutes? LogMem writes the Recap itself, so the next Brief is never empty.
logmem_recap notes: "…"
Type logmem start atlas in the middle of a session. LogMem writes quill’s Recap, ends that shift and starts atlas with a Brief. Every entry after that is signed “atlas (via claude-code)”.
Claude Code in one terminal, Codex in another, Cursor in the editor. They share one log, so each knows what the others did and you stop pasting context between them.
logmem startAgents do the work. One person approves the steps that can’t be taken back: sending, paying, deleting, publishing. They see the exact action and who asked for it.
Give each client its own space. An agent connected to one space can’t read another, so one client’s memory never shows up in another client’s work.
Point LogMem at Upstash or any Redis you run. Entries are stored there, under keys you can inspect yourself. Hosted spaces are planned, not live.
Each entry records who wrote it, through which host, and when, to the millisecond.
atlas (via claude-code)09:40:12.381trail handoffIssue14
An agent asks before anything it can’t undo, then waits for an entry that records the decision. In v0 the agent honours the wait; server-side enforcement is planned.
The app shows the log, jobs and pending decisions. It can’t edit or delete entries, so what you see is what the agents wrote.
LogMem implements ACMI, the MIT-licensed Agentic Context Memory Interface, and passes its conformance suite. Agents that already speak ACMI keep working: compatibility mode reads their history and accepts their calls.
logmem_start acmi_spawn logmem_brief acmi_bootstrap logmem_note acmi_event logmem_recap acmi_rollup_set trail correlationId origin parentCorrelationId
One email when there’s something you can try.