-
Notifications
You must be signed in to change notification settings - Fork 0
Core concepts
GenOS borrows language from version control, event sourcing, evolutionary systems, and causal experimentation. This glossary explains how those ideas fit together.
The genome is the declarative blueprint of an agent: identity, role, drives, objectives, tool permissions, and memory policies. Genomes are versioned and immutable. A mutation creates a descendant genome with an explicit parent and delta instead of rewriting its ancestor.
The phenotype is observed behavior produced by a genome in a particular environment and history. Two agents with the same genome can diverge because their worlds, events, or experience differ. GenOS experiments can record that divergence rather than treating the prompt as the only cause.
Drives represent tunable behavioral tendencies such as exploration, risk tolerance, or verification threshold. Traits can be measured, associated with lineage, and promoted only when evidence supports the claim.
State is the mutable cognitive context: working memory, beliefs, goals, event cursor, and runtime metadata. It is separated from the genome so temporary knowledge does not silently become inherited identity.
A belief or claim can carry confidence, evidence, contradictions, origin, and lineage. GenOS includes experiments for belief updates, contradiction detection, replication, heritability, and promotion.
An event is an append-only record of a meaningful transition: initialization, tool execution, state change, branch creation, checkpoint, failure, merge, or other supported operation. Events form the input to replay and the basis of an audit trail.
Sibling events can share a correlation identifier while remaining partitioned by branch. Causation-chain examples model how an earlier event contributed to a later outcome; they do not claim universal causal inference from arbitrary logs.
A world is the environment assigned to a capsule. Implemented providers include logical and directory-backed worlds, plus Git-worktree-oriented workflows. A world has a root, identity, snapshots, and containment rules.
A snapshot is an immutable checkpoint over supported agent and world state. It links to ancestry and integrity data and can be restored, forked, diffed, or replayed.
Content-addressed objects are identified by a digest of their contents. This supports integrity checks and deduplication: identical immutable content can be stored once even when referenced by multiple snapshots.
The Agent–World Capsule is the runnable boundary combining genome, state, world, history, and budget. A capsule has a lifecycle such as created, running, paused, completed, failed, or budget-exhausted.
A fork creates one or more descendant capsules from a common baseline. Each child receives its own branch identity, mutable state, event partition, and—when applicable—isolated world.
A hypothesis records the reason for a branch. It turns parallel execution into an experiment: branches are not merely “attempt 1” and “attempt 2”; they test named alternatives against comparable evidence.
Counterfactual replay returns to a checkpoint, changes an intervention or assumption, and compares the resulting trajectory with the historical one. It is useful for incident analysis and “what if?” experiments.
A GenOS diff can compare more than text. Depending on the workflow it may report file-tree changes, state deltas, goals, beliefs, scores, or lineage.
Evaluation converts branch outcomes into structured evidence: pass/fail gates, scores, objective vectors, receipts, and reviewer decisions.
When objectives conflict, a branch is dominated if another is at least as good on every objective and better on at least one. Non-dominated branches form the Pareto frontier and remain candidates for policy or human selection.
Promotion makes a verified branch the accepted successor. In the safest workflow, the branch is restored or replayed from the baseline and checked again before promotion.
A cognitive merge synthesizes selected evidence or compatible state from branches under a manifest. It is evidence-aware reconciliation, not a blind concatenation of memories.
Snapshots and genomes retain parent relationships, producing an acyclic provenance graph. Lineage queries answer which checkpoint, mutation, branch, or merge produced a result.
An experience packet is a reviewed, typed unit of reusable knowledge derived from a branch. It can be synthesized into another context without copying the branch's complete internal history.
Mutation creates a child genome from one parent and an explicit delta. Breeding records two-parent lineage and child validation. Both are research mechanisms for controlled agent variation—not autonomous self-modification without policy.
Branch evolution allocates a global compute budget, terminates weak branches, gives survivors additional budget, and can recursively split promising candidates. The current implementation provides the mechanics; live evaluator integration remains incomplete.
Untrusted inputs and tool outputs can be marked tainted until a verification step accepts them. Taint is metadata supporting policy; it is not by itself a complete information-flow security system.
A circuit breaker stops or cools down repeated failures and runaway loops. Together with budgets and branch termination, it bounds experimental execution.
Isolation is always scoped. Directory containment and branch-local event streams are implemented boundaries. Arbitrary process, network, kernel, credential, and remote-service isolation require additional controls outside the current local runtime.
GenOS is active pre-alpha research software. Verify maturity and evidence before relying on a capability. · Repository · License · Security
Start
Inside the system
- Core concepts
- Architecture
- Isolation, replay & provenance
- Hallucination reduction
- Evaluation, evolution & memory
- Orchestrator & 77 strategies
- Organizations, teams & swarms
Use GenOS
See it in action