The open-source reliability layer for AI agent memory — contradiction detection, health scoring, and consolidation that wraps your existing vector backend.
Benchmark — Track 2 · Behavioral
7 real-world contradiction scenarios across 3 systems, scored on correctness, conflict signal, and fact preservation. All three return the right answer. Only memnotary tells you when another answer is wrong.
6/7
contradiction scenarios flagged
2.0×
conflict detection rate vs Mem0
0.94
overall score · highest tested
Signal measures whether conflicting or superseded memories are flagged in results. Correctness is 1.00 for all three systems — the difference is whether the wrong answer is silently returned alongside the right one.
Read the full benchmarkLive demo
store → detect → score → consolidate. No API key, no external service — reproduced entirely in a terminal with an in-memory adapter.
How it works
01
One constructor. Works with Qdrant, Chroma, pgvector, or in-memory. Nothing in your storage layer changes.
from memnotary import (
Memnotary,
ContradictionDetector,
QdrantAdapter,
)
from qdrant_client import QdrantClient
mn = Memnotary(
QdrantAdapter(QdrantClient(url="localhost:6333")),
detector=ContradictionDetector(llm_fn=your_llm),
)02
When a new memory lands, memnotary searches for semantic neighbors above its similarity threshold. Confirmed conflicts become ConflictRecords — inspectable, queryable, actionable.
await mn.store(Memory(agent_id="bot",
text="Refund policy is 30 days", embedding=...))
await mn.store(Memory(agent_id="bot",
text="Refund policy changed to 14 days", embedding=...))03
health() returns a snapshot across your full memory collection. consolidate() plans and executes: supersede, merge, or flag for human review.
h = await mn.health("bot")
# contradiction_score: 1.00 risk: HIGH
await mn.consolidate("bot")
h = await mn.health("bot")
# contradiction_score: 0.00 risk: LOWThree calls. Your agent memory goes from a silent liability to an auditable, self-healing store. Get started →