SESSION WORKING SETSessions swap in or out for this Evaluation
FULL / METABuilt for durable work and concurrent execution
One Agent. Many Objectives. Advancing in parallel.
Morphz lets one Agent maintain multiple Sessions, Objectives, and Threads inside one Context. Conversation does not wait for long-running work, cognitive changes commit through transactions, and execution can safely reach different target nodes.
One Context carries dialogue, goals, and execution.
OBJECTIVE · THREADDurable goals and causal work advance separately
DURABLEEXECUTION TARGETWork routes to a node with the required capabilities
SCOPEDContext Transaction
The Agent proposes cognitive change; the Runtime validates version, authority, causality, and commit boundaries.
derive · revise · retire · restoreCore capabilities
Let an Agent maintain its Context, advance concurrent work, and safely reach real environments.
Context Encoding, explicit transactions, durable scheduling, and Execution Targets work together as the foundation for long-running Agent work.A self-maintaining Context. Cognition that keeps evolving.
Morphz organizes long-term memory as persistent, versioned, evaluable cognitive state. The Runtime reports physical facts and capacity pressure; the Agent continuously revises its cognition through explicit transactions.
Session is part of Context
One Context Encoding organizes Inbox, Mind, Session Directory, Kernel state, and the sole evaluation entry in one recursive structure.
No automatic lossy Context compaction
Under capacity pressure, the Runtime reports the boundary. The Agent explicitly decides what to preserve, revise, or retire instead of letting the system silently rewrite the entire Context as a summary.
Current and recent Sessions enter Context
The current Session comes first. By default, Morphz projects up to 50 Sessions active within the last 24 hours; the rest remain durable and can return when active again.
Cognition evolves through evidence
derive, revise, retire, restore, and protect produce cognitive changes with explicit versions, provenance, and lifecycles.
One Agent. Many workstreams in progress.
Morphz uses Sessions, Objectives, Threads, and Activations for distinct interaction, goal, causal, and execution semantics. Tools and long-running work do not occupy the whole Agent.
Keep talking while the Agent works
Tool execution, long-running work, and new messages can proceed concurrently. Results return along their original causal routes without blocking other Sessions.
Durable goals converge independently
Each Objective owns state, dependencies, and completion conditions across Evaluations, waits, and process restarts.
Causal work and execution are separate
A Thread preserves the causal path; an Activation represents one execution opportunity and can resume from a defined boundary after a wait.
Concurrent updates pass version checks
Independent cognitive changes can commit. Stale or conflicting changes must be evaluated again instead of overwriting newer state.
Run tasks on target nodes. Keep authority in the Runtime.
Execution Targets separate cognitive evaluation from physical execution. Tasks can run locally, on a managed SSH host, or on an Edge Node within the target environment's sandbox and minimum authority.
Local, Managed SSH, and Edge Node
One Agent can route different tasks to where resources live while retaining explicit target ownership.
Native isolation on macOS, Linux, and Windows
Workspace writes are enforced by the operating-system sandbox. Restricted mode fails closed when its required isolation backend is unavailable.
Minimum authority, exact scope, revocable access
Capability leases bind Principal, Agent, Context, Thread, Target, and operation set. Approval does not implicitly widen authority.
Credentials stay out of Context; execution remains traceable
Secrets stay in a separate Secret Store. Approvals, tool calls, and state transitions enter the event record.
04 / COGNITIVE APPLICATIONS
Package how an Agent works as a cognitive application.
A Cognitive Application organizes domain capabilities. A Harness defines how an Evaluation runs. Yao gives the model and Runtime a shared cognitive evaluation language. Applications can replace the working method without bypassing Runtime transactions, authority, or execution boundaries.Explore the Morphz open standards→Cognitive Application
Organizes interfaces, domain resources, tools, and integrations for a specific field of work.
Harness
A versioned package defines the Evaluation Loop, domain contract, default Mind, and policy.
Yao
Typed cognitive evaluation programs support infer, eval, and par, with Runtime admission before execution.
05 / DEFINITION
Multiplexed Session I/OFrame · Session swap in / swap outThread · Objective schedulingCapability-governed execution
beyond the single conversation loop
A new operating model for agents.
Morphz is an open-source agent for long-running, concurrent work. In one continuously running Runtime, it unifies multiplexed Session I/O, Frame and Session working-set swapping, Thread and Objective scheduling, and capability-governed execution. Together, these mechanisms form its cognitive operating-system core.06 / FIRST RUN
Run Morphz on your own machine.
Install a prebuilt release with one command, then connect a model service through Setup and start Morphz.Run Morphz→curl -fsSL https://github.com/morphz-ai/morphz/releases/latest/download/install.sh | sh07 / LEARN MORE
From product capabilities to implementation.
Read the product docs, technical article, and research paper, or inspect the source and run Morphz locally.ARTICLEStructured Context EvaluationAn introduction to the Morphz computational model.↗PAPER · EVALUATIONResearch paper and evaluationUpdated STATE-Bench-derived protocol: Morphz 122/150, Letta 93/150, Mem0 96/150.↗SOURCESource and testsThe Apache-2.0 open-source implementation.↗DOCSProduct documentationInstallation, configuration, concepts, and operation.↗