my-project-public

repository

loading code, commits, and activity

repositories

loading repo index

#1	# Clawd Memory SOTA Architecture - Masterplan
#2	## Temporal Epistemic Graphs with Veracity-Weighted Consolidation
#3
#4	This masterplan now targets Clawd Memory: the local-first agent brain for Clawd. The Clawd layer owns the agent contract, markdown vault, Solana/OODA metadata, and recall behavior. The Mnemosyne engine remains the SQLite storage and retrieval substrate until a deliberate migration is planned.
#5
#6	---
#7
#8	## Paper Citations & Foundations
#9
#10	### Core Papers
#11	1. Memanto (arXiv:2604.22085) - Typed Semantic Memory with Information-Theoretic Retrieval
#12	- 13-type schema, zero ingestion cost, single-query retrieval
#13	- Authors: Moorcheh AI / EdgeAI Innovations
#14
#15	2. Moorcheh ITS (arXiv:2601.11557) - From HNSW to Information-Theoretic Binarization
#16	- 32x compression, deterministic exhaustive scan, 9.6ms latency
#17	- Authors: Abtahi et al., Ontario Tech University
#18
#19	3. REMem (ICLR 2026, arXiv:2602.13530) - Reasoning with Episodic Memory
#20	- Hybrid memory graph, time-aware gists, agentic retrieval
#21	- Authors: Shu et al., Ohio State / Intuit AI
#22
#23	4. HippoRAG (arXiv:2405.14831) - Neurobiologically Inspired Long-Term Memory
#24	- Hippocampal indexing theory, graph-based retrieval
#25	- Authors: Various
#26
#27	### Supporting Papers
#28	5. BEAM Benchmark - Beyond a Million Tokens (2026)
#29	6. Hindsight Blog (2026/04) - SOTA on BEAM with structured memory
#30	7. Honcho Benchmarks (blog.plasticlabs.ai) - User modeling and dreaming
#31
#32	---
#33
#34	## Our Novel Contribution
#35
#36	"Temporal Epistemic Graphs with Veracity-Weighted Consolidation"
#37
#38	No existing system combines:
#39	- Typed semantic memory (Memanto)
#40	- Binary vector compression (Moorcheh)
#41	- Episodic graph structure (REMem)
#42	- Veracity-weighted confidence (our invention)
#43	- Deterministic retrieval (our constraint)
#44	- Zero LLM ingestion (our constraint)
#45
#46	---
#47
#48	## Implementation Phases
#49
#50	### Phase 0: Research & Foundation (COMPLETE)
#51	- [x] Identify all relevant papers
#52	- [x] Map techniques to our architecture
#53	- [x] Define 10 core principles
#54	- [x] Create masterplan document
#55
#56	### Phase 1: Typed Memory Schema
#57	Goal: Implement 13-type deterministic classification
#58	Papers: Memanto (arXiv:2604.22085)
#59	Innovation: Rule-based typing with zero LLM calls
#60
#61	Types:
#62	1. fact - Objective, verifiable info
#63	2. preference - User/system preferences
#64	3. decision - Choices affecting future
#65	4. commitment - Promises/obligations
#66	5. goal - Objectives to achieve
#67	6. event - Historical occurrences
#68	7. instruction - Rules/guidelines
#69	8. relationship - Entity connections
#70	9. context - Situational info
#71	10. learning - Lessons from experience
#72	11. observation - Patterns noticed
#73	12. error - Mistakes to avoid
#74	13. artifact - Document/code references
#75
#76	Implementation:
#77	- Pattern-based classifier using regex + keyword matching
#78	- Confidence scoring based on pattern match strength
#79	- No LLM calls during classification
#80	- Store type in dedicated column
#81
#82	### Phase 2: Information-Theoretic Binary Vectors
#83	Goal: Replace float32 vectors with binary compression
#84	Papers: Moorcheh ITS (arXiv:2601.11557)
#85	Innovation: Deterministic exhaustive scan over binary vectors
#86
#87	Implementation:
#88	- Maximally Informative Binarization (MIB) algorithm
#89	- Efficient Distance Metric (EDM) using Hamming distance
#90	- Information-Theoretic Score (ITS) for ranking
#91	- 32x memory reduction
#92	- No HNSW index needed
#93	- SQLite-native storage
#94
#95	### Phase 3: Episodic Gist+Fact Graph
#96	Goal: Build hybrid memory graph with temporal edges
#97	Papers: REMem (arXiv:2602.13530)
#98	Innovation: Time-aware gists + structured facts in single graph
#99
#100	Implementation:
#101	- Gist extraction: concise episode summaries with timestamps
#102	- Fact extraction: (subject, predicate, object) triples
#103	- Graph nodes: V_gist ∪ V_phrase
#104	- Graph edges: E_rel ∪ E_ctx ∪ E_syn
#105	- Temporal qualifiers: point_in_time, start_time, end_time
#106
#107	### Phase 4: Veracity-Weighted Consolidation
#108	Goal: Bayesian confidence scoring + conflict resolution
#109	Papers: None (our invention)
#110	Innovation: Confidence-based synthesis with automatic conflict detection
#111
#112	Implementation:
#113	- Veracity tiers: stated (1.0), inferred (0.7), tool (0.5), imported (0.6), unknown (0.8)
#114	- Bayesian updating: confidence = 1 - (0.7^n) where n = mention count
#115	- Conflict detection: contradictory facts with same subject+predicate
#116	- Resolution: higher confidence wins, lower confidence flagged
#117	- Consolidation: periodic background synthesis of high-confidence facts
#118
#119	### Phase 5: Polyphonic Recall Engine
#120	Goal: Multi-strategy parallel retrieval with deterministic re-ranking
#121	Papers: Memanto + REMem
#122	Innovation: Parallel strategies with cross-strategy confirmation boost
#123
#124	Implementation:
#125	- Strategy 1: Binary vector similarity (Moorcheh)
#126	- Strategy 2: FTS5 keyword search
#127	- Strategy 3: Temporal index lookup
#128	- Strategy 4: Entity graph traversal
#129	- Strategy 5: Fact triple matching
#130	- Re-ranker: weighted combination with cross-strategy boost
#131	- No LLM calls during retrieval
#132
#133	### Phase 6: Integration & Testing
#134	Goal: Full system integration with comprehensive testing
#135	Tests:
#136	- Unit tests for each component
#137	- Integration tests for end-to-end flow
#138	- BEAM benchmark validation
#139	- Ablation studies (disable each component, measure impact)
#140	- Performance benchmarks (latency, throughput, memory)
#141
#142	### Phase 7: Paper Draft
#143	Goal: arXiv preprint with full methodology and results
#144	Sections:
#145	1. Introduction + Related Work
#146	2. Core Principles
#147	3. Architecture
#148	4. Implementation Details
#149	5. Benchmark Results
#150	6. Ablation Studies
#151	7. Cost Analysis
#152	8. Conclusion
#153
#154	---
#155
#156	## Success Metrics
#157
#158	\| Metric \| Current \| Target \| SOTA \|
#159	\|--------\|---------\|--------\|------\|
#160	\| BEAM 100K \| 21.8% \| 40%+ \| 73.4% \|
#161	\| Ingestion Latency \| ~500ms \| <10ms \| N/A \|
#162	\| Query Latency \| ~100ms \| <50ms \| N/A \|
#163	\| Memory Overhead \| 1x \| 0.03x (32x comp) \| N/A \|
#164	\| LLM Calls per Ingest \| 1+ \| 0 \| N/A \|
#165	\| LLM Calls per Query \| 1+ \| 0 \| N/A \|
#166
#167	---
#168
#169	## Risk Mitigation
#170
#171	1. Binary vector quality loss: Test on BEAM before/after, tune MIB threshold
#172	2. Type classification accuracy: Validate against labeled dataset, adjust patterns
#173	3. Graph traversal performance: Add depth limits, cache frequent paths
#174	4. Consolidation correctness: Conservative thresholds, human review for high-stakes
#175
#176	---
#177
#178	## Timeline
#179
#180	\| Phase \| Duration \| Cumulative \|
#181	\|-------\|----------\|------------\|
#182	\| Phase 1: Typed Schema \| 2 days \| 2 days \|
#183	\| Phase 2: Binary Vectors \| 3 days \| 5 days \|
#184	\| Phase 3: Episodic Graph \| 4 days \| 9 days \|
#185	\| Phase 4: Veracity Consolidation \| 3 days \| 12 days \|
#186	\| Phase 5: Polyphonic Recall \| 3 days \| 15 days \|
#187	\| Phase 6: Testing \| 5 days \| 20 days \|
#188	\| Phase 7: Paper \| 5 days \| 25 days \|
#189
#190	Total: ~25 days to SOTA paper
#191
#192	---
#193
#194	## Implementation Order
#195
#196	Starting with Phase 1 (Typed Schema) because:
#197	1. Highest impact on BEAM scores (ABS, SUM, KU categories)
#198	2. Lowest risk (rule-based, no dependencies)
#199	3. Foundation for all other phases
#200	4. Can be tested independently
#201
#202	Then Phase 2 (Binary Vectors) for performance gains.
#203	Then Phase 3 (Episodic Graph) for reasoning improvements.
#204	Then Phase 4 (Veracity) for consolidation quality.
#205	Then Phase 5 (Polyphonic Recall) to tie it all together.
#206
#207	Ready to start Phase 1?
#208

z6Mkq5mY3JWtxoxUobWcfNHm7AkRubgSWEZTkBVqZXJviFZ5/my-project-public