LONG HORIZON AGENTS HACK · VERIFIED LIVE RUN

Forget intelligently.
Wake at the right moment.

SPORE moves conditionally useful evidence out of an agent’s active context, watches the live web, and restores the evidence only when the world changes.

See the autonomous runRead the code ↗
RUN VERIFIEDLoading proof…
THE PROBLEM

Agents either remember too much or forget what becomes useful later.

Keeping every rejected idea in model context wastes tokens and attention. Deleting it loses the reasoning needed when a price, API, regulation, or capability changes. SPORE stores the detailed evidence outside working context and leaves behind a compact, typed wake condition.

ONE GOAL · ZERO STAGE CLICKS

The agent ran the whole loop itself.

AUTONOMOUS RUN VERIFIED
LIVE EVIDENCE

Every claim stays inspectable.

Liquid planned and interpreted. Nimble returned live URLs. A source policy accepted only citations from the official OpenAI domain family.

SPONSOR TOOLS · WORKING PATH

Each tool owns a necessary job.

01

Liquid AI

Creates the search plan, analyzes evidence, and decides the post-wake action with schema-constrained local inference.

PLANNER · ANALYST · DECISION02

Nimble

Runs multiple current web searches and returns source URLs for the watcher’s evidence trail.

LIVE WEB RESEARCH
03

SQLite + archive

Persists memory state, removes active context transactionally, and protects detailed evidence with SHA-256.

CONDITIONAL MEMORY
04

RawTree

Receives the lifecycle event batch and proves delivery by returning every event ID for the completed run.

REMOTE AUDIT TRAIL
WHY IT COUNTS

The scheduler starts itself.

One user command supplies the goal. After that, the agent plans its research, creates and sleeps the memory, starts AutonomousWatcher.start(), evaluates current sources, wakes exactly once, rehydrates context, and acts.