Naive memory
retracted citationThe therapy trial showed sustained benefit across patients in all cohorts.
10.5555/retracted-example
Belief-aware memory at runtime
Most agents wake up with amnesia. Ours wakes up knowing what it must unlearn.
GroundTruth is a belief-maintenance memory layer built on Cognee — beliefs carry evidence, get contested, are superseded, and can be surgically forgotten.
Why this exists
Memory only grows. Facts rot. Your agent keeps citing them.
Memory only grows
Long-running agents accumulate stale facts by default.
Facts rot
Retractions, CVEs, and reversals happen faster than rebuild cycles.
Your agent keeps citing them
Downstream answers stay stale until the memory is actively corrected.
How it works
The 4-verb cycle
remember
Evidence in
Ingest claim data and source evidence into grounded memory.
groundtruth/ingest.py
recall
Evidence-annotated answers
Answering surfaces belief states and evidence classes.
groundtruth/answer.py
improve
Contested to adjudicated
Session + semantic evidence flow into resolve or revision actions.
groundtruth/feedback.py, groundtruth/contest.py
forget
Surgical removal
Session/authoritative revocations retire or supersede stale facts.
groundtruth/watcher.py
Evidence ladder
Where belief changes come from
- authority feed: deterministic updates from external signal
- semantic inference: conflict and alignment scoring
- user assertion: adjudication and explicit feedback
Benchmark headline
Proven committed numbers
Naive
18/20
retracted originals cited
GroundTruth
0/20
retracted originals cited
Controls
5/5
control retention
V2 semantic
28/28
coverage; precision 1.00 / recall 0.67
Live demo pages
Explore the product
Ask
Side-by-side naïve and belief-maintained answers plus retraction control.
Briefing
Learned, contested, revised, and purged belief queues.
Contested
Review conflicts and resolve them into explicit state transitions.
Timeline
Time-travel diff for what changed and why.
Evidence
The benchmark sheet and belief-maintenance metrics in full.