01
Planning
Form a search strategy, revise it after new evidence, and keep the next action deliberate.
plan · analyzeText
Context management for long-horizon agents
Keep the evidence.
Drop the clutter.
ContextPilot gives LLM agents a disciplined way to plan, retrieve, remember, and offload—so they can navigate long documents while keeping their working context focused, compact, and evidence-rich.
- Structured memory
- Targeted retrieval
- Soft offloading
Working contextKeep what matters.
5.6k tokens
Active contextcompact
ANALYZE✦PLAN✦SEARCH✦READ✦REMEMBER✦OFFLOAD✦ANSWER
A context control plane
Four moves.
One compact mind.
ContextPilot turns context management into explicit, inspectable actions. The model can gather evidence without carrying every intermediate result all the way to the answer.
02
Retrieval
Index once, search precisely, then open only the chunks needed to verify the answer.
searchEngine · readChunk
03
Memory
Save durable facts with their relationships and retrieve them when final reasoning needs them.
memorize · updateMemory
04
Offloading
Delete noise, truncate bulky observations, or compress useful spans into a smaller working set.
delete · truncate · compress
Interactive execution trace
Explore a live trace.
Step through an illustrative local trace. No model endpoint, document upload, or API key is used in this browser demo.
USER QUERY
Active messages2 in context
Trained where choices matter
Explore key actions.
Assign fine-grained credit.
The RL pipeline focuses exploration on context-editing decisions and uses downstream outcomes to train the intermediate states that produced them.
01 · Context-aware partial rollout
Explore high-impact actions.
Branch around pivotal context edits instead of exploring every action uniformly across the trajectory.
02 · Fine-grained credit assignment
Assign fine-grained credit.
Propagate downstream rewards to the context snapshots that shaped each outcome along the trajectory.
Evaluation suite
Four ways to go long.
Run the full pipeline across long-book comprehension, narrative QA, conversational memory, and deep web search—or launch one task at a time.
01InfBenchLong-book choice
02NovelQANarrative reasoning
03LongMemEvalConversational memory
04BrowseComp+Deep search
# Install the inference environment
bash infer/scripts/setup_environment.sh
source infer/.venv/bin/activate
# Evaluate all four tasks
bash infer/scripts/run_full_pipeline.sh \
/path/to/checkpoint my-runOpen source