Key Insights: Memory Provider Landscape¶
The Central Problem: Context vs. RAG¶
The user correctly identifies that most "memory" systems are just RAG for old conversations. True context preservation requires: 1. Understanding — not just storing text but extracting meaning 2. Synthesis — deriving insights across multiple interactions 3. User modeling — building a model of who the user is 4. Contextual recall — surfacing the RIGHT memory at the RIGHT time 5. Temporal awareness — knowing when facts change over time
Best Ideas from Each Provider (for Super-Memory-TS vNext)¶
From Honcho¶
- Dialectic reasoning: LLM-synthesized insights about the user, not just raw facts
- Multi-peer profiles: Separate profiles per agent persona
- Cold/warm prompt selection: Different reasoning depth based on context availability
- Orthogonal config knobs: Separate controls for cost, depth, and frequency
From OpenViking¶
- Tiered context loading (L0/L1/L2): Abstract → Overview → Detail
- Filesystem hierarchy: Structured, browsable knowledge organization
- 80-90% token reduction: Load only what's needed
From Mem0¶
- Server-side LLM extraction: Automatic fact extraction without manual curation
- Dual memory scope: Session + User memory layers
- Circuit breaker pattern: Graceful degradation when memory fails
- Apache 2.0 license: Commercial-friendly
From Hindsight¶
- Knowledge graph: Structured facts with entity relationships
- Reflect synthesis: Cross-memory synthesis deriving higher-level insights
- Multi-strategy retrieval: Temporal + Entity + Metadata + BM25 in parallel
- Best benchmarks: 91.4-94.6% on LongMemEval
From Holographic¶
- Trust scoring: Asymmetric feedback (penalize wrong more than reward right)
- Contradiction detection: Auto-detect conflicting memories
- HRR algebra: Compositional queries (AND across entities)
- Zero dependencies: SQLite-only, works anywhere
From RetainDB¶
- Full chronological retrieval: Complete timeline, not lossy semantic search
- Turn-by-turn extraction: Atomic memory processing
- Hybrid search: Vector + BM25 + reranking
From ByteRover¶
- Human-readable knowledge tree: Markdown files, editable, inspectable
- Pre-compression extraction: Capture before context window squeezes
- Curation engine: ADD/UPDATE/UPSERT/MERGE/DELETE operations
- Best LoCoMo score: 92.2%
From Supermemory (the cloud service, not the user's project)¶
- Context fencing: Prevent recursive memory pollution
- Memory relationships: Update/Extend/Derive graph connections
- Session-end graph ingest: Build knowledge graph from conversations
- Multi-container mode: Isolate memories per project/context
Architecture Gaps in Super-Memory-TS¶
What it does well:¶
- Local-first, privacy-preserving
- Fast vector search (HNSW, <10ms)
- Project indexing with semantic chunking
- MCP-native (OpenCode compatible)
- Tiered search strategies (TIERED/PARALLEL)
- Project isolation
What's missing for "true context preservation":¶
- Automatic memory extraction — currently requires manual
add_memorycalls - Knowledge graph — no entity relationships or memory connections
- User modeling — no persistent profile of the user
- Memory synthesis — no cross-memory reflection/insight generation
- Tiered context loading — no L0/L1/L2 abstraction levels
- Trust scoring — all memories weighted equally
- Contradiction detection — conflicting memories coexist silently
- Temporal awareness — no concept of facts changing over time
- Memory decay/consolidation — old memories never compress or fade
- Context fencing — no protection against memory pollution
- Pre-compression hooks — doesn't capture before context squeeze
- Multi-modal memory types — no distinction between facts, preferences, decisions, patterns
Recommendation: The "Super-Memory" vNext Architecture¶
A hybrid approach combining the best ideas: 1. Keep: MCP protocol, local-first, Qdrant/HNSW, project indexing 2. Add: Automatic LLM-based extraction layer (inspired by Mem0/Honcho) 3. Add: Lightweight knowledge graph with memory relationships (inspired by Hindsight/Supermemory) 4. Add: Tiered context loading L0/L1/L2 (inspired by OpenViking) 5. Add: Trust scoring + contradiction detection (inspired by Holographic) 6. Add: Cross-memory synthesis/reflect pass (inspired by Hindsight) 7. Add: User profile modeling (inspired by Honcho) 8. Add: Context fencing + pre-compression hooks (inspired by Supermemory/ByteRover) 9. Add: Memory types/categories with different decay rates (inspired by Mem0/RetainDB) 10. Add: Human-readable export format (inspired by ByteRover)