Skip to content

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":

  1. Automatic memory extraction — currently requires manual add_memory calls
  2. Knowledge graph — no entity relationships or memory connections
  3. User modeling — no persistent profile of the user
  4. Memory synthesis — no cross-memory reflection/insight generation
  5. Tiered context loading — no L0/L1/L2 abstraction levels
  6. Trust scoring — all memories weighted equally
  7. Contradiction detection — conflicting memories coexist silently
  8. Temporal awareness — no concept of facts changing over time
  9. Memory decay/consolidation — old memories never compress or fade
  10. Context fencing — no protection against memory pollution
  11. Pre-compression hooks — doesn't capture before context squeeze
  12. 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)