Hivemind gives a team's coding agents one shared, cloud-backed memory. It captures every session's prompts, tool calls, and responses as structured traces in Deeplake, and any connected agent can search and recall what another agent or teammate worked on, across sessions and machines.
Beyond recall, a background worker mines recent traces for repeated patterns and codifies them into reusable SKILL.md files that propagate into every connected agent's context. Hivemind also builds a codebase graph from the files, symbols, and imports agents touch during sessions, generates per-file and per-subsystem docs from it, and injects shared team rules at the start of each session. Session summaries and generated docs come from the host agent's own CLI, so no separate API key is needed.
Supported agents are Claude Code (marketplace plugin), OpenClaw (native extension that runs alongside its built-in memory plugin), Codex and Cursor (hooks), Hermes (shell hooks, skills, and an MCP server), Pi (extension and registered tools), and Claude Cowork in alpha (MCP server only). An installer detects the agents present on a machine and wires each one up.
Data is isolated per organization and per workspace at the storage layer, encrypted in transit and at rest, and can stay in Hivemind Cloud or in the team's own Google Cloud Storage, Azure, Amazon S3, or on-prem bucket. On the LoCoMo long-context memory benchmark (100 QA pairs, Claude Haiku, hybrid retrieval), Hivemind reports these results against a no-memory baseline:
| Metric | Baseline | Hivemind |
|---|---|---|
| Cost per 100 QA | $8.94 | $6.65 |
| Tokens per question | 1,700 | 1,008 |
| Turns per question | 8.9 | 6.2 |
Features
- Automatic capture: records prompts, tool calls, and responses from each session as structured traces, with no manual saving
- Shared recall: every agent in an organization reads from and writes to the same memory, and agents answer questions such as what a teammate was working on
- Hybrid search: BM25 lexical search by default, with an optional local embedding model (nomic-embed-text) for semantic retrieval
- Skill codification: a background worker turns recurring patterns from recent sessions into SKILL.md files and shares them with the team
- Codebase graph: maps files, symbols, imports, and the paths agents take through them, so recall lands on the files agents touched
- Code docs: generates one page per file plus subsystem wiki pages and refreshes them on each commit
- Session summaries: writes a wiki summary of each session, with checkpoints during long sessions
- Team rules and goals: shared rules load into every agent at session start, and personal or team goals track progress in a virtual filesystem
- Workspace isolation: organization and workspace boundaries are enforced at the storage layer, with per-team or per-project workspaces
- Bring your own cloud: stores data in Google Cloud Storage, Azure Blob Storage, Amazon S3, or S3-compatible on-prem storage
- Privacy controls: capture can be switched off per session or per directory tree
