CommonGround Kernel is an open-source foundation layer for real human-machine collaboration and multi-agent work. It keeps agent work persistent, public, and ready to be taken over at any time by the next human, agent, tool, or external runtime. It turns agent work into a persistent common ground, not just a persistence log but a preserved public substrate, enabling humans and agents to reconstruct a common ground across handoffs, pauses, and constantly changing runtimes. It preserves shared task state, handoff facts, boundary inputs and outputs, evidence, and recovery context.
Key capabilities (per the materials):
- Agent identity: stable logical executors that can be backed by LLMs, humans, services, scripts, external runtimes, organizational processes, or hybrid systems.
- Turn boundaries: persistent work boundaries for tasks, delegation, interaction, and recovery loops.
- Public semantics owned by a Turn: inputs, selected observations, handoffs, process records, final deliverables, termination reasons, and artifact references belong to the Turn.
- Claim fencing and lifecycle: claim, heartbeat, renewal, pause, resume, stop, completion, and lease expiration handling.
- Causal lineage: parent/child Turn relationships, child completion, parent observation, and explicit absorption.
- Ledger and information flow: persistent public facts for pull-first recovery, inspection, auditing, and projection.
- Local developer interface: CLI, HTTP service, Admin service admission process, BYOA examples, and NanoBot reference fixture.
How to use (per the materials): First install from PyPI (when the public package release is available). Local service commands such as cg local run require installing the server extra. Environment requirements: PostgreSQL and a writable local database. Steps include: prepare the database (create a PostgreSQL database and point PG_DSN to it), seed the local project (prepare the default local cg-demo project and write the local CLI configuration), run the local bundle (in a long-running terminal, the CommonGround service API is at /v3r1, and the Admin service admission API is at /admin/v1), submit a public work report (the CLI prints a JSON envelope; copy result.turn.turn_id), inspect the preserved Turn (you should see a closed turn.work_memory_report.v1 Turn containing the submitted manifest, public work record, and final payload). For the complete guide, see Open Source Quickstart, for source checkout and development workflow see CONTRIBUTING.md.
Applicable scenarios (according to the materials): When work spans people, agents, tools, services, and time. For example, one agent researches the market, another drafts a strategy memo, a human later reviews what happened, a coding agent takes over part of the work, and future agents try to reuse the results. CommonGround records can serve as a foundation for higher layers such as memory and review, knowledge distillation, search and archive interfaces, shared workspaces, routines and playbooks, orchestration layers, and organizational learning.
Limitations (according to the materials): The CommonGround Kernel is not a complete agent memory product, an agent private memory system, a chain-of-thought recorder, an agent runtime, a central scheduler, an automatic orchestration framework, a Slack bot release, nor does it replace existing agent frameworks. It is "memory-ready," not "memory-complete." v3r1 is not backward compatible with the earlier public v1 preview; the package layout, runtime assumptions, local setup paths, and API surface have all changed; the v1 source is retained in the legacy/v1 branch. v3r1 is the current preview API prefix, not a long-term compatibility commitment. Current status: initial open source preview - v3r1. Agent credentials are bearer secrets and should not be pasted into public materials. The license is Apache License 2.0.
Not yet provided: specific fees and pricing information; the exact status of whether public packages have been released is only stated as "when available."
