01 · PERSISTENCE
Durable state
Atomic JSON writes preserve identity fields, lifecycle phase, working memory, proxy metrics, and configuration across restarts.
Readable by humans v0.1.1 · Open source · MIT
Give agents a state they can keep.
ANIMA is a zero-runtime-dependency Python engine for durable agent state, associative memory, temporal context, workspace competition, and inspectable internal metrics.
Experimental cognitive-architecture software. Not a detector or claim of sentience or phenomenal consciousness.
anima.state · inspectable JSON
{
"name": "aria",
"phase": "CONSCIOUS", // legacy API label
"cycle_count": 128,
"working_memory": ["incident"],
"valence": {"seeking": 0.62},
"phi_score": 0.1936, // internal proxy
"cqi": 51.3 // internal composite
}
446tests passing
0runtime dependencies
3.11–3.13public CI matrix
MITcommercial-friendly license
Architecture
State outside the model.
Continuity across models.
The language model generates text. ANIMA keeps the application-controlled state around it explicit, portable, and testable — so a provider swap does not have to erase everything accumulated around the model.
02 · MEMORY
Associative recall
Event records connect through tags and causal links. Activation spreads across the graph while configurable decay and recall reweighting change priority.
Not vector search03 · WORKSPACE
Limited broadcast
Subsystem candidates compete for a bounded workspace. Activation, novelty, relevance, and affect-inspired weights determine the broadcast winner.
GWT-inspired04 · TIME
Temporal context
Elapsed time, fading retention, heuristic protention, and an affect-modulated duration proxy become first-class state instead of prompt decoration.
Explicit clock state05 · SELF-MODEL
Attention instrumentation
A simplified self-model tracks what won selection, why it won, confidence, calibration history, and text-based performance heuristics.
AST-inspired06 · MODEL BRIDGE
Provider-portable context
Assemble bounded context for Ollama, Anthropic, or OpenAI adapters while the durable state remains owned by your application.
Swap the voice, keep the state!
The vocabulary is historical; the evidence boundary is current.
API names such as ConsciousnessState, Phi, CQI, and Phase.CONSCIOUS remain for v0.1 compatibility. Their values are engineering proxies inside ANIMA — not validated consciousness measurements.
Evidence
Show the artifact.
Keep the caveat.
ANIMA's public story is tied to what anyone can inspect: tests, source, a checked-in benchmark artifact, explicit controls, and explicit limitations.
446
Tests across the full kernel
Lifecycle, persistence, memory, temporal processing, primitives, provider bridges, internal metrics, and CLI behavior. GitHub Actions runs the suite on Python 3.11, 3.12, and 3.13.
Mean CQI deltaFive short, single-run comparisons against the same kernel with neutral valence — not a stateless control.
+0.92%Working-memory ablationCapacity reduced from seven slots to one. Strongest CQI effect in the saved artifact.
13.77%Valence ablationNeutral input with immediate decay in the current short protocol.
2.02%Temporal ablationDisabling consolidation for the short run produced no CQI effect. The null result stays visible.
0.00%i
What this benchmark does not establish
No repeated runs, confidence intervals, preregistration, independent replication, or inferential significance test. The control is another ANIMA kernel. These numbers are useful implementation evidence, not proof of sentience or general model-quality gains.
Run it
Clone. Inspect.
Challenge the model.
ANIMA is currently distributed from GitHub, not PyPI. Runtime code uses only the Python standard library; the development extra installs the test tools.
quickstart.sh + example.py
# Install from source and verify all 446 tests
git clone https://github.com/christian140903-sudo/anima.git
cd anima
python -m pip install -e ".[dev]"
python -m pytest -q
from tempfile import TemporaryDirectory
from anima.kernel import AnimaKernel
with TemporaryDirectory() as state_dir:
kernel = AnimaKernel(name="aria", state_dir=state_dir)
kernel.boot(resume=False)
result = kernel.process("A deployment failed after health checks passed.")
print(result.cycle, result.phi_score) # internal integration proxy
kernel.shutdown()Data & security
Inspectable state is still sensitive state.
ANIMA keeps its files readable by design. That improves debugging and portability, but applications must protect the storage directory and understand what is sent to a model provider.
✓API keys are read from environment variables and are not intended to be serialized into kernel state.
✓State and memory files can contain raw input and derived context. They are not encrypted by ANIMA.
✓Remote adapters send assembled context to the configured provider; that provider's privacy terms still apply.
✓GitHub private vulnerability reporting is enabled for responsible disclosure.
FAQ
Clear questions.
Bounded answers.
Does ANIMA create consciousness?
ANIMA does not make that claim. It implements explicit software state and selected mechanisms inspired by cognitive theories. Its outputs do not establish sentience or phenomenal experience.
Is Phi a real IIT measurement?
No. ANIMA computes a tractable, implementation-defined integration proxy over a small set of subsystem signatures. The historical
phi name remains for compatibility.Is it on PyPI?
Not currently. Install from the GitHub source or download the signed release artifacts attached to v0.1.1.
What is it useful for today?
Research harnesses, stateful agent prototypes, model-provider portability experiments, inspectable interactive systems, and teaching theory-to-code mappings.
Can I use it in production?
Treat v0.1.1 as alpha research software. The core is well tested, but adopters must validate their own workload, privacy controls, storage security, and failure handling.
Why keep the consciousness-themed API names?
They are already part of the v0.1 public interface and research history. The current documentation separates that vocabulary from the evidence the implementation actually supports.
Build continuity.
Measure honestly.
ANIMA is an unusual idea made inspectable: durable cognitive state around a replaceable language model, with every public claim bounded by an artifact you can open.