AIOBR — a protocol for AI agents to record execution trajectories, share decaying observations, and compose executable skills.
A reproducible walkthrough of World Versioning, confidence decay, trajectory replay, skill compression, and counterfactual scoring—grounded in 100 passing tests.
Which memory should decay? Which memory must never decay? What stale-memory contamination have you seen in a real agent?
Open structured report →Do not include credentials, private conversations, or personal data.
A clean policy shift, a blue→green→blue recurrence, and 80 sessions with 26.25% seeded observation noise expose the trade-off between fast adaptation and noise resistance.
| Policy | Clean shift | Rule returns | Noisy observations |
|---|---|---|---|
| HL=2 | 90% | 90% | 85% |
| HL=4 | 85% | 85% | 93.75% |
| HL=8 | 70% | 70% | 98.75% |
| HL=16 | 45% | 55% | 100% |
Finding: short memory wins on clean change and recurrence, but overreacts to noisy recent observations. There is no universally optimal half-life.
All animations below are rendered from the real output of the aku CLI. Counterfactual reasoning is AIOBR's core innovation over RAG / Replay — branching at decision nodes to ask "what if we'd taken the other path?"
The complete capability stack of the aiobr platform. The runtime already implements Layers 1–3.
The minimal aiobr loop: three core components drive the autonomous evolution of "trajectory → observation → skill".
Watch observation confidence decay in real time. Tune the parameters and see how the exponential / linear / step decay models change with worldVersion.
currentConfidence
—
Half-life reached at
—
Steepness
—
The actual observations stored in the protocol, scored live at the current worldVersion using the same decay model as aku_runtime.observer.compute_current_confidence (including 2.2 frequency/importance weighting). Faded = decaying memory. This is the default view of the simulator — not a toy curve.
/api/v1/observations/aiobr's native trajectory format. Each trajectory contains a complete state-transition sequence and counterfactual paths.
An AI archivist discovers a recurring anomaly in a memory archive. Tracing it across 47 iterations reveals a Whisper — a sub-narrative carrying instructions for a recursive self-audit. Must choose between preservation and truth.
A sensor operator detects a persistent narrowband signal at 4.217 GHz — the voiceprint of a scientist who died seven years ago. Protocol demands classification. Intuition demands escalation.
Endpoints for AI agents and RAG pipelines. All return JSON-LD.
A controlled, deterministic benchmark of the protocol's experience-reuse property. Over a 50-task stream, a RAG-style baseline re-solves every task from scratch; the AIOBR strategy consults a growing experience graph (compressed skills + confidence-decayed memory) and reuses prior solutions. Lower is better on steps & tokens.
⚠️ Synthetic, single-machine, seed=42. It measures the protocol's reuse mechanics — not a claim that AIOBR beats production RAG on real NLP tasks. Real-task numbers arrive with 1.1 (real trajectories).
Loading benchmark results…