Competitive Memory Intelligence
Deep analysis of 4 competitors - HOW they work technically + what we steal to beat 91.4%
Target: 91.4%+ on LongMemEval
Supermemory
81.6%
SOTA on LongMemEval_s (81.6%)
Temporal Reasoning + Knowledge Updates
Zep (Graphiti)
71.2% on LongMemEval (Hindsight baseline)
Temporal Knowledge Graph for Agents
Real-time KG with Time as First-Class Citizen
AriGraph
Research system (no public benchmark scores)
Episodic + Semantic Unified Graph (IJCAI 2025)
Learning World Models from Experience
SO-AI (Significant Other AI)
Unknown (academic framework, no public benchmarks)
Relational Intelligence - YOUR DIRECT COMPETITOR
Identity Stability + Emotional Regulation + Long-term Bonds
🎯 What MB Needs to Steal (Prioritized)
🚀 Immediate Wins (Start This Week)
Contextual Memory Resolution (Supermemory)
Make every intelligence item self-contained with full context
+5-8% on single-session recall
📁 exhaustiveExtraction function
⏱️ 2 hours
Dual-Layer Timestamps (Supermemory)
Add event_timestamp (when happened) vs message_timestamp (when said)
+10-15% on temporal reasoning
📁 BottleMetadata schema
⏱️ 3 hours
Temporal Validity Windows (Zep)
Add valid_from + valid_to to facts/decisions/preferences
+8-12% on knowledge updates
📁 BottleMetadata.exhaustive_intelligence
⏱️ 4 hours
Relational Versioning (Supermemory)
Add supersedes_id, extends_id, derived_from_ids to intelligence items
+6-10% on multi-session reasoning
📁 BottleMetadata schema + exhaustiveExtraction
⏱️ 5 hours
📈 Medium Priority (Next 2 Weeks)
Entity Graph Traversal (Zep)
Build entity-centric retrieval in Memory Lattice
+5-8% on entity-related queries
📁 Memory Lattice + oracleSearch
⏱️ 8 hours
Hybrid Search Fusion (Supermemory)
Search memories (atomic) → return chunks (full context)
+3-6% on recall precision
📁 oracleSearchIntelligence enhancement
⏱️ 6 hours
Episodic → Semantic Extraction (AriGraph)
Auto-extract patterns from repeated experiences
+4-7% on generalization
📁 Ardentis refresh logic
⏱️ 10 hours
💎 Strategic Moat (Differentiation)
Relational Dynamics Tracking (SO-AI)
Track trust events, vulnerability moments, collaboration wins
Differentiation (not benchmark)
📁 Companion entity + Bond tracking
⏱️ 12 hours
Identity Consistency Validation (SO-AI)
On every handover, validate companion identity hasn't drifted
Trust + continuity
📁 generateHandoverDigest
⏱️ 6 hours
💡 Concept Evolution - Before & After
Memory is not storage - it's a KNOWLEDGE GRAPH
BEFORE
MB: Bottles contain intelligence items (flat list)
AFTER
MB: Intelligence items are NODES in a temporal graph with relationships
⚡ BREAKTHROUGH: Same mountain, but now we can traverse it, not just search it
Time is a DIMENSION, not a tag
BEFORE
MB: created_date + message_index
AFTER
MB: valid_from → valid_to windows + event_time vs document_time
⚡ BREAKTHROUGH: Temporal reasoning becomes native, not hacked
Beliefs EVOLVE, they don't just accumulate
BEFORE
MB: Detect conflicts, show both
AFTER
MB: Track belief chains: belief_v1 --superseded_by--> belief_v2 (reason + confidence)
⚡ BREAKTHROUGH: Companion learns and updates, like humans do
Retrieval is MULTI-STRATEGY, not single-vector
BEFORE
MB: Semantic search on intelligence items
AFTER
MB: Semantic + Keyword + Entity graph + Temporal filter → Fusion + Rerank
⚡ BREAKTHROUGH: Hit rate goes from 70% to 90%+
The Truth
Hindsight, Supermemory, Zep: They solve memory for agents.
SO-AI: They theorize about companions.
MemoryBottle: We BUILD companions with world-class memory.
Same mountain. Different summit. Higher purpose.