shared reasoning memory for multi-agent systems
When multiple AI agents research the same problem independently, they re-derive the same knowledge from scratch. Agent 2 doesn't know what Agent 1 already found. The result: redundant LLM calls, wasted tokens, and contradictory conclusions that no one catches.
hivememory gives agents a shared memory layer. Agents store structured reasoning artifacts, query what others have found before doing their own work, and get automatic alerts when findings contradict each other.
Agents store structured claims with evidence, confidence scores, and provenance links. Not raw text. Each artifact records who produced it, what supports it, and what it builds on.
When a new artifact is stored, FAISS finds semantically similar claims. If two artifacts are close but have divergent confidence, a conflict is flagged. An optional LLM check confirms true contradictions.
Every artifact records its dependencies as artifact IDs, forming a directed acyclic graph. The DAG answers "which agent's work did this build on?" and enables cascading invalidation.
from hivememory import HiveMemory, Evidence
hive = HiveMemory()
# store a finding
art = hive.write(
claim="Voice AI market projected to reach $50B by 2028",
evidence=[Evidence(source="industry report", content="35% CAGR", reliability=0.9)],
confidence=0.85,
agent_id="researcher-1",
)
# query shared memory before doing new research
existing = hive.query("voice AI market size", top_k=3)
# check for contradictions
open_conflicts = hive.get_conflicts()
============================================================ CONFLICT DETECTION DEMO ============================================================ --- Agent 1 (optimistic analyst) storing findings --- stored: The AI code editor market is projected to reach $5B by 2026 stored: GitHub Copilot holds 55% market share among AI code editors stored: AI code editors improve developer productivity by 40-55% stored: Cursor has an NPS score of 72 among professional developers Artifacts in memory: 4 Conflicts so far: 0 --- Agent 2 (conservative analyst) storing findings --- CONFLICT: AI code editor market estimated at $2.1B in 2026 → High similarity (0.96) but divergent confidence: 0.55 vs 0.95 CONFLICT: Copilot's market share has declined to 35% → High similarity (0.72) but divergent confidence: 0.55 vs 0.90 CONFLICT: AI code editors improve productivity by 15-25% → High similarity (0.96) but divergent confidence: 0.60 vs 0.95 stored: Enterprise adoption of AI code editors reached 78% in 2025 ============================================================ PIPELINE RESULTS ============================================================ Artifact pairs compared: 22 Stage 1 candidates (sim > 0.7): 5 Stage 2 confirmed contradictions: 3