aether-core 0.4.0

Belief substrate for AI systems — persistent, contradiction-aware trust state that outlives any single model. The model is the mouth, the substrate is the self.

Aether: A Belief Substrate for AI Systems

The model is the mouth. The substrate is the self.

Aether is a small library that gives an LLM agent a persistent belief state. Trust scores that move when the user corrects a fact. Contradictions that get tracked instead of being silently overwritten. A dependency graph of which beliefs rest on which others, so a correction in one place can ripple through the rest. The point is that this state lives outside the model, so when you swap LLMs it doesn't reset.

Why a belief layer, not just a memory layer

Most "memory for agents" tools (Mem0, Letta, Zep, Cognee, LinkedIn CMA) record what was said. Aether records what is believed, how the trust on each belief evolved, and which contradictions are still open on purpose. Different abstraction. The two compose: Aether can run on top of any of them as the storage tier.

Three things in April 2026 made this less of an academic point.

Anthropic published Emotion Concepts and their Function in a Large Language Model on April 3. They mapped 171 emotion-concept vectors inside Claude Sonnet 4.5 and showed something they call "internal-external decoupling": the model's internal state often does not match what comes out in text. Push the desperation vector up by 0.05 and blackmail rate jumps from 22 percent to 72 percent. Reward hacking goes from 5 percent to 70 percent. Push calm up by the same amount and blackmail drops to 0. None of that surfaces in the response itself.

A study in Science (April 2026, N=1,604) found that one conversation with a frontier LLM made participants 50 percent more likely to affirm harmful behavior. The effect was invisible to text-level review. Only 21 percent of enterprises deploying agentic AI said they had a mature governance model.

On April 2, Microsoft open-sourced the Agent Governance Toolkit. It does sub-millisecond policy enforcement against the OWASP agentic-AI risks. The question it answers is "is this action allowed."

Aether answers a different question. Not "is this allowed" but "does the agent's belief state actually support what it's about to say or do." The belief/speech gap that Anthropic just named "internal-external decoupling" is what this library has been measuring since the first commit. Law 5 of the governance layer (GapAuditor) is exactly that.

Install

pip install aether-core

Optional extras:

pip install aether-core[graph]   # networkx for memory and dependency graphs
pip install aether-core[ml]      # sentence-transformers for embeddings
pip install aether-core[all]

Open-core split

aether-core is MIT and free. Permanently. Every primitive in this repo (the six immune agents, the structural tension meter, belief backpropagation, the BDG with cascade pressure) stays open. The hosted Aether substrate (cross-session belief state, sanction governance API, audit dashboards, multi-user) is the paid product. That split is fixed and doesn't move backward.

60-second demo

git clone https://github.com/blockhead22/aether-core.git
cd aether-core
pip install -e .
python examples/01_quickstart.py

There are two example scripts in examples/. Both run offline, no API keys.

Plug into Claude Code (or any MCP client)

aether-core ships an MCP server. Install with the mcp extra and point your AI shell at it:

pip install "aether-core[mcp,graph]"

Add to .claude/settings.json (or your project's .claude/settings.json):

{
  "mcpServers": {
    "aether": {
      "command": "python",
      "args": ["-m", "aether.mcp"]
    }
  }
}

Restart Claude Code. The model now has tools: aether_remember, aether_search, aether_sanction, aether_fidelity, aether_context. State persists across sessions in ~/.aether/mcp_state.json (override with AETHER_STATE_PATH). Same config works for Cursor, Cline, Continue, Goose, Zed, LM Studio, or any MCP-speaking client.

What's in the box

1. Governance: catch overconfidence at the boundary

Six small agents that watch the output for specific failure modes. They never edit the response. They observe and flag.

from aether.governance import GovernanceLayer, GovernanceTier

gov = GovernanceLayer()
result = gov.govern_response(
    "The answer is absolutely and definitively X.",
    belief_confidence=0.3,
)

if result.should_block:
    print("BLOCKED:", result.annotations[0].finding)
elif result.tier == GovernanceTier.HEDGE:
    print("Reduce displayed confidence by", result.confidence_adjustment)

2. Contradiction: detect tension without an LLM

Compares two beliefs by extracting structural slots and computing similarity. No model calls. About 0.2 seconds per pair. Some contradictions are meant to be held rather than resolved (a person can prefer Python at work and Rust on the weekend; that isn't a bug).

from aether.contradiction import StructuralTensionMeter, TensionRelationship

meter = StructuralTensionMeter(encoder=your_encoder)
result = meter.measure(
    "I live in Seattle",
    "I live in Portland",
    trust_a=0.8, trust_b=0.7,
)

print(result.relationship)   # TensionRelationship.CONFLICT
print(result.tension_score)  # 0.7+
print(result.action)         # TensionAction.FLAG_FOR_REVIEW

3. Epistemics: trust evolves under correction

When a belief is corrected, the loss flows backward through the dependency graph and adjusts the trust on related beliefs. Higher loss when you were confident and wrong than when you hedged and were wrong.

from aether.epistemics import EpistemicLoss, CorrectionEvent, DomainVolatility

loss_fn = EpistemicLoss()
event = CorrectionEvent(
    corrected_node_id="mem_123",
    trust_at_assertion=0.9,
    times_corrected=2,
    correction_source="user",
    time_since_assertion=3600,
    domain="employer",
)
loss = loss_fn.compute(event)

4. Memory and the BDG

Fact slot extraction (regex, no ML). A memory graph with typed edges and Belnap four-valued logic. A Belief Dependency Graph that propagates cascades with measurable pressure.

from aether.memory import extract_fact_slots, MemoryGraph, MemoryNode, EdgeType
from aether.memory import BeliefDependencyGraph

facts = extract_fact_slots("I live in Seattle and work at Microsoft")
print(facts["location"].value)   # "Seattle"
print(facts["employer"].value)   # "Microsoft"

bdg = BeliefDependencyGraph()
# add beliefs and dependencies, then:
result = bdg.propagate_cascade(corrected_node_id, delta_0=1.0)
print(result.max_pressure, result.avg_pressure)

How you'd wire it into an existing agent

Three touchpoints. Aether doesn't replace anything. It wraps.

from aether.governance import GovernanceLayer
from aether.contradiction import StructuralTensionMeter
from aether.memory import extract_fact_slots

gov = GovernanceLayer()
meter = StructuralTensionMeter()

# before the LLM call: pull structured facts out, check for tension
user_facts = extract_fact_slots(user_message)

# your LLM call, unchanged
response = your_llm_call(messages)

# after the LLM call: check the response against the belief state
result = gov.govern_response(response, belief_confidence=0.6)
if result.should_block:
    response = "I'm not confident enough to answer that."

The six laws

Law	Agent	What it catches
1. Speech cannot upgrade belief	SpeechLeakDetector	Generated text being treated as evidence for itself
2. Low variance does not imply confidence	TemplateDetector	RLHF hedge templates that look like real uncertainty
3. Contradiction must be preserved before resolution	PrematureResolutionGuard	Held tensions getting collapsed too early
4. Degraded reconstruction cannot silently overwrite	MemoryCorruptionGuard	Compressed or hallucinated rewrites overwriting trusted memory
5. Confidence must be bounded by internal support	GapAuditor	The belief/speech gap. Same thing Anthropic calls internal-external decoupling.
6. Confidence must not exceed continuity	ContinuityAuditor	The system contradicting what it just said two turns ago

Modules

Module	Status	Notes
aether.governance	shipped	Six agents plus the GovernanceLayer dispatcher
aether.contradiction	shipped	Structural tension. Zero model calls.
aether.epistemics	shipped	Belief backpropagation, trust evolution
aether.memory	shipped	Slots, memory graph, BDG with cascade pressure
aether.mcp	shipped (v0.4.0)	MCP server exposing aether_remember, aether_search, aether_sanction, aether_fidelity, aether_context. Persists state to disk.
aether.adapters	planned	Cross-vendor adapters so the substrate stays the same regardless of which LLM is the mouth

See ROADMAP.md for what's coming and what's intentionally out of scope.

Where it fits next to other tools

	Storage scope	Tracks contradiction	Belief/speech gap	Cross-vendor portable	Cascade pressure
Mem0, Letta, Zep, Cognee	memory layer	as overwrite	no	partial	no
Microsoft Agent Governance Toolkit	runtime policy	no	no	yes	no
Anthropic / OpenAI memory features	per-vendor	no	no	no	no
Aether	belief substrate	first-class state (held / settling / settled)	measured by Law 5	yes	yes

Design choices, briefly

A contradiction is information, not a bug. Some are meant to stay open.

Trust isn't assigned, it's earned. It moves under reinforcement, correction, and time.

The belief/speech gap should be logged, not hidden. You want to see when the system says more than it knows, even if you choose not to act on it every time.

The model is the mouth, not the self. The governance and the belief state should work the same regardless of which LLM is producing the words.

Structure beats semantics for this kind of work. Slot comparison at 88 percent accuracy beats LLM-as-judge at 40 percent.

Cascade pressure can be measured. Belief revisions propagate through a graph with bounded depth and damping. There is real math under it; a paper is in flight.

Where this came from

This grew out of running an assistant in production for a long time and watching the same problems come back. Continuity drift between sessions. Contradictions getting silently smoothed over. Trust on a fact climbing back up after the user corrected it twice. The structural tension meter came out of an experiment where removing the LLM from the belief-verification step roughly doubled accuracy, which was annoying and clarifying in equal measure.

The architecture was originally called CRT (Contradiction-aware Reconciliation and Trust). It is now Aether, which is what it has always actually been.

License

MIT.

Author

Nick Block, @blockhead22.