Commonplace

Research on knowledge systems, running as one.

Commonplace is a living doctrine for agent-operated knowledge systems, developed and tested by running one. The doctrine selects and coordinates how model-mediated and symbolic operations are used. Explicit artifacts can activate model capabilities and give their use project authority; code and validators can faithfully execute operations that should not be reconstructed on every call. The doctrine, prompts, code, and models can all change. Like the Ship of Theseus, Commonplace remains the same project through a governed sequence of revisions, not because any component is permanent.

Its first application is an LLM wiki, in the sense AI researcher Andrej Karpathy sketched: a persistent, linked Markdown layer around a person's or project's work. Human-directed agents turn vague thoughts into retained notes, connect them to evidence and related claims, and revise both the knowledge base and its operating machinery. Humans direct the inquiry and remain responsible for judgments that current evidence and evaluation cannot settle.

The repository is Commonplace's current embodiment. It contains adopted doctrine, research and evidence that can challenge it, and the procedures and code that make it operative. Research does not become doctrine merely by being stored here.

Two separate pressures keep the design modular. Different collections support different kinds of work, so task-specific types and link conventions stay local. Structures can also become obsolete as questions, evidence, or model capabilities change, so local choices remain revisable rather than accumulating by default. Commonplace keeps shared invariants small for both reasons. See why task-fitted structure costs cross-task reuse and why a framework rule with a boundary-preserving rival is not an inherited constraint.

Threads worth following

The automated software house conjecture. The current research program asks whether a complete software house can sustain open-ended coherent change with current learned components pinned and no human in an internal production role. Its starting project theory and production machinery may be written by people. The test is whether the resulting composite can carry the program-theory function, use it where the needed implication was not stated verbatim, revise when later demands expose an error, and continue across a declared scope and horizon. How such a house should acquire and improve its organization is a separate question. The Naur note reopens the bearer question, while the coherent-search note states the longitudinal test. The software house as the unit of training says how such a house should learn, and bootstrapping the first automated software house says how it would be built from a house that still has people inside.

Deployment-time learning. Durable changes to behavior-shaping prompts, rules, tools, schemas, tests, and code can affect later sessions without updating model weights. Storage is insufficient: later operation must load or enforce the result. Deploy-time learning develops this path, while bounded context explains why selective routing remains necessary. The learning theory index maps the wider thread.

Self-improving systems. A system improves itself only when evidence-responsive change reaches its own behavior-determining organization. One architecture directly updates behavior; another searches candidates, evaluates them, and retains an accepted proposal. Reflection is a separate property that provides addressability, not improvement by itself. The self-improving systems index and Commonplace case develop the distinction.

Agent-usable memory. Agents need discoverable, composable, and trusted knowledge under bounded context. Explicit link semantics expose support, contrast, and consequence; routing selects what reaches a task; provenance and review indicate how strongly to rely on it. Because information value is observer-relative, useful condensation depends on the consuming agent and task.

Systems compared. We reviewed 148 agent memory systems, including Mem0, Graphiti, Cognee, and Letta. The comparative analysis finds that activation and verification distinguish the reviewed systems more than storage location. The collection over-samples file-based systems, so its counts describe this corpus rather than the field.

Browse

Use it yourself

Commonplace is open source. You can:

  • Vendor the knowledge base read-only inside a project so agents can consult the research and external-system reviews.
  • Install the system to give a project the types, conventions, skills, and commands for building its own agent-operated knowledge base.

The GitHub README covers the tool, INSTALL.md gives the setup flow, and the installation architecture explains the design.