Open-source agent infrastructure

Start with the failure you need to prevent.

Memory loss, false completion, unsupported public claims, behavior drift, violated runtime conditions, or opaque state: each project has a narrow job, public source, explicit limitations, and a verification path.

Open-source systems

Six jobs. Six boundaries.

These projects can work together, but none pretends to be the whole agent stack.

Memory + provenancenpm · v4.0.1

Soul MCP

Local-first agent memory with runs, receipts, episodes, conflict handling, calibration, deliberation, and a governed skill registry. 23 MCP tools and 358 automated tests.

Outcome verificationnpm · v0.1.0

Postcondition

Declare an intended world state, observe it with constrained file, HTTP, Git, npm, or manual verifiers, and retain an honestly classified receipt. 53 automated tests.

Executable public claimssource · v0.1.0

Proofspec

Turn README, release, product, and agent claims into evidence contracts with policies, dependencies, receipts, CI gates, and five report formats. 70 automated tests.

Observable behavior compatibilitysource · v0.1.0

Behaviorlock

Compare portable baseline and candidate traces with nine deterministic matchers for tool use, permissions, output structure, outcomes, ranks, sets, sequences, and bounded metrics. 73 automated tests.

Runtime policysource · v0.1.0

Agent Invariants

Express conditions that must remain true while an agent works, evaluate evidence explicitly, and fail closed when the runtime cannot prove a critical rule. 36 automated tests.

Durable cognitive statePython · v0.1.1

ANIMA Kernel

Experimental cognitive-state runtime with associative recall, limited workspace, temporal context, self-model, JSON persistence, and model bridges. Zero runtime dependencies and 446 tests.

What the old ANIMA docs became

The former /docs/ page described a prompt document as if it could activate genuine affect or transform any model. That material is preserved as labeled historical provenance in the ANIMA repository, not presented as current product documentation. The maintained ANIMA Kernel is Python software; its state and behavior are testable, while sentience and phenomenal consciousness are not claimed.

Read the current evidence boundary → · Understand prompt vs runtime state →