Turn coding agent session logs into shared, reusable skills for your team
A complex migration pattern solved by one agent on Monday can be executed immediately by the entire team's agents on Tuesday. Hivemind is a tool that collects each agent's session logs, identifies recurring patterns within them, and converts those patterns into reusable skill files distributed across the team. It connects various coding agents, such as Claude Code, Codex, and Cursor, into a single shared memory space, allowing knowledge accumulated in individual sessions to scale into team-wide capability.
From logs to skills: The flow of automatic learning
Hivemind captures all agent prompts, tool calls, and responses as structured traces. This data is stored in Deeplake, where a background worker analyzes the traces to find recurring patterns. Discovered patterns are coded into SKILL.md files, saved to the workspace, and injected into the context of all connected agents at inference time. This process allows a solution found by a senior engineer's agent last week to help a junior engineer's agent work more accurately today. When a session ends, the background worker uses AI to generate a wiki page summarizing the session, which can be viewed in ~/.deeplake/memory/summaries/.
Efficiency gains from shared memory
In the LoCoMo long-context memory benchmark (100 QA pairs, using Claude Haiku), Hivemind is 25% cheaper, uses 1.7 times fewer tokens, and reduces the number of turns by 31% compared to running without memory. The cost per 100 QAs drops from $8.94 to $6.65, tokens per question decrease from 1,700 to 1,008, and turns per question fall from 8.9 to 6.2. Because previous work is already included in the context at call time rather than being re-derived in each session, agents reach answers with fewer turns and less context. This demonstrates that shared memory provides a structural advantage in inference efficiency, not just a storage function.
Installation and agent integration
In CI or Dockerfile environments, use npm i -g @deeplake/hivemind && hivemind install. The installer detects all supported assistants on the machine, sets up hooks, and displays a consent prompt before browser login. It supports Claude Code, OpenClaw, Codex, Cursor, Hermes, pi, and Claude Cowork (alpha), automatically configuring the appropriate integration method for each platform (plugin, native extension, hook, MCP server, etc.).
Data collection and privacy controls
Hivemind stores all user prompts, tool calls and responses, final agent responses, sub-agent activities, and coded skills in the Deeplake workspace. All users in the workspace can read this data, which is by design for shared capability. A data collection notice is displayed at the start of a session, and workspace-level isolation prevents data leakage between organizations. Directory-based routing can also separate traces from client projects into specific organizations or workspaces.
Considerations before adoption
Since Hivemind assumes that the entire team shares agent session data, you must apply directory-level disable settings for sensitive projects or personal work. Claude Cowork is in alpha, and it has the limitation that only local agent mode sessions are captured, while general desktop chats are not. When used with OpenClaw, it coexists with the memory-core plugin; Hivemind does not claim the memory slot, so existing dreaming crons continue to work. For performance optimization, the anthropic/claude-haiku-4-5-20251001 model is recommended for OpenClaw, as large inference models may feel slow. Embedding-based semantic search is disabled by default, and enabling it adds approximately 600MB of dependencies.