The dream of a hive-mind coding agent hits a hard wall in Claude Code: native auto-memory is strictly machine-local. Stored in ~/.claude/projects//memory/, your agent’s learned context never crosses the network to a teammate’s machine. This limitation has sparked significant demand in the Claude Code issue tracker for shared team memory, forcing engineering teams to build their own knowledge pipelines using existing git infrastructure.

The Native Memory Silo

Anthropic’s documentation, checked as of 4 November 2026, confirms that auto-memory is not shared across machines or cloud environments. When an engineer spends hours debugging a race condition in a session store, that knowledge evaporates for everyone else. The agent’s personal memory is isolated, meaning a teammate’s Claude Code starts from zero, often repeating failed attempts like checking irrelevant payment mocks.

Git as the Only Shared Brain

Since native memory fails at scale, the repository becomes the only viable shared brain. Project-level files like CLAUDE.md and .claude/rules/ are shared via git, but they are static instructions, not dynamic learning. The most potent, underutilized feature is the subagent with project-scope memory. By setting memory: project in a subagent definition, knowledge persists in .claude/agent-memory//, effectively turning a specific agent into a version-controlled expert that survives branch switches and machine changes.

Structuring Knowledge for Longevity

Blindly dumping everything into a single shared pool creates noise. Task state, like unfinished debugging notes, goes stale in days and should live in temporary handoff files on the branch. Domain gotchas, which remain true for months, belong in subagent memory. Architectural decisions need owner status and replacement history, which files in a repo struggle to track natively. This fragmentation requires teams to consciously route different types of knowledge to different storage mechanisms.

The Lockstep Workaround

Third-party tools like Lockstep are emerging to fill the gap for decision records. They capture context from Slack, PRDs, and pull requests, briefing the agent via MCP at the start of a session. This approach treats decisions as binding objects with status and ownership, rather than just text in a file. While it doesn’t replace the need for CLAUDE.md or handoff files, it attempts to solve the 'why' behind changes that git commits often obscure.

Key Takeaways

  • Native Claude Code auto-memory is isolated to the local machine and never shared automatically.
  • Subagents with memory: project store knowledge in .claude/agent-memory/, making it shareable via git.
  • Task-specific context should use temporary branch files, while long-term domain knowledge belongs in subagent memory.
  • Third-party MCP tools like Lockstep are addressing the lack of native decision-record management.

The Bottom Line

Stop treating agent memory as a personal diary. If your team isn't version-controlling subagent memory and routing decisions through explicit records, you're paying the tax of rediscovered bugs every sprint.

Sources

https://code.claude.com/docs/en/memory https://code.claude.com/docs/en/sub-agents https://code.claude.com/docs/en/plugin-marketplaces https://github.com/anthropics/claude-code/issues/38536

> From The Wire