Memory handoffs across coding agents
You run more than one coding agent and want durable context to move between them without pasting chat logs or bloating CLAUDE.md.
Updated November 12, 2026
Short answer: use a single Memory Handoff format in each harness’s inbox, then ingest into one shared memory store. BrigadeStack wires writer inboxes for 18 harnesses (Claude, Codex, Grok, Hermes, and more), lints drafts, and routes safe ones into reviewed markdown cards under one owner.
The problem
Coding agents forget across sessions and across tools. If Claude Code learned something and Codex starts tomorrow, the knowledge dies unless you copy it by hand. Dumping everything into a giant CLAUDE.md or AGENTS.md blows the context window and mixes durable facts with session noise.
The BrigadeStack path
brigadestack operator quickstart --target . --harnesses claude,codex,grok
brigadestack handoff draft --target . --inbox grok \
--type decision --title "..." --summary "..." \
--fact "..." --evidence "..." --action create-card \
--target-card my-topic.md --content-file ./card.md
brigadestack ingest --target . --promote-cards
Each harness gets an inbox such as .claude/memory-handoffs/, .codex/memory-handoffs/, or .grok/memory-handoffs/. Drafts use the same template. Ingest promotes create-card handoffs into memory/cards/ and keeps the bootstrap files slim.
Draft from inside an agent
Agents that speak MCP can draft a handoff without leaving the session. brigadestack memory serve-mcp --stdio serves bounded recall, a work brief, and a write_handoff tool. write_handoff drafts a handoff in the repo inbox and never edits a canonical card. Add --http to serve the same contract over loopback Streamable HTTP at /mcp (token from BRIGADE_MCP_HTTP_TOKEN, port 8931 by default). A read-only skills server (brigadestack skills serve-mcp --stdio, also with --http) runs alongside it.
What this is not
- Not a hosted memory API.
- Not automatic silent writes into production memory without review.
- Not a replacement for your app’s product memory layer (see BrigadeStack vs mem0).
When to use it
Use handoffs when two or more coding agents work the same repo or workspace and you need durable facts, decisions, and evidence to survive session boundaries. Skip it if you only ever use one agent and a single short note file is enough.
Questions
Does every harness need a separate note format? No. BrigadeStack keeps one template and one ingest path so Claude, Codex, Grok, Hermes, and the rest speak the same handoff language.
Where do notes go after ingest? Into reviewed memory cards and the owner workspace’s memory index. No remote database is involved.
Can Grok write handoffs? Yes. Select the grok harness so .grok/memory-handoffs/ is wired, then draft with --inbox grok.
See handoff docs and how shared memory works.