Intelligence Layer
The Intelligence Layer uses LLMs and advanced algorithms to keep your memory store healthy, your retrieval sharp, and your agents onboarded faster.Extraction
Turns raw text (conversations, logs, documents) into structured memory operations using LLMs. The extractor classifies each piece of information as an ADD, UPDATE, DELETE, or NOOP operation.Auto Entity & Relationship Extraction
Automatically extracts entities (services, people, tools, infrastructure) and relationships (depends_on, uses, manages, monitors) from decision traces and memory entries using LLMs. Extracted data is stored in dedicated tables with confidence scores, temporal validity, and trace provenance. Enable withAMFS_AUTO_EXTRACT=true.
Memory Critic
Automated quality analyzer that scans the memory store and detects five issue classes:- Toxic — repeated negative correlations
- Stale — entries not referenced in a long time
- Contradictory — conflicting entries for the same key
- Uncalibrated — confidence scores misaligned with outcome history
- Orphaned — entries with no links to any outcome or other entries
Memory Distiller
Compacts large stores into smaller, higher-quality sets via pruning, consolidation, and bootstrap set generation for agent onboarding.Memory Safety Validator
Pre-write guardrails: contradiction detection, temporal consistency, confidence thresholds, and causal chain integrity.Multi-Strategy Retrieval
Combines semantic, BM25, temporal, and confidence signals via Reciprocal Rank Fusion (RRF). When a learned ranking model is trained, it automatically receives 30% weight in the fusion pipeline.Pro MCP Tools
| Tool | Description |
|---|---|
amfs_critique | Run the Memory Critic and get a quality report |
amfs_distill | Trigger distillation (prune, consolidate, or generate bootstrap set) |
amfs_validate | Validate a candidate memory before writing |
amfs_retrain | Train the learned ranking model from outcome data |
amfs_calibrate | Learn optimal confidence multipliers from outcome history |
amfs_export_training_data | Export decision traces as SFT/DPO/reward model datasets |
amfs_record_llm_call | Record an LLM call with model, tokens, cost, and latency |
amfs_graph_path | Find shortest trust-weighted path between two entities in the knowledge graph |
amfs_graph_query | Flexible graph edge search by relation, entity type, or confidence range |