Skip to main content

​
Intelligence Layer

The Intelligence Layer is the set of LLM-driven and statistical tools that keep what your agents have learned healthy, make retrieval sharp, and let a new agent start from the best of an existing one. Most of it lives on the Pro Tools page of the dashboard and behind the Pro MCP tools listed at the end.

​
Entity & Relationship Extraction

Reads raw text — conversations, logs, documents, decision traces, and memory entries — and pulls out the entities in it (services, people, tools, infrastructure) along with how they relate to each other. An LLM does the extraction; the results are stored with confidence scores, temporal validity, and a link back to the trace they came from, which is what populates the knowledge graph. Six relation types are recognised: Extraction runs on demand: open Entity Extraction in the dashboard and run it over recent traces, or call it through the API. Extracted entities and relations are reviewable there before you rely on them in the graph.

​
Memory Critic

Automated quality analyser that scans everything learned 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 a large store into a smaller, higher-quality one. The amfs_distill tool generates a bootstrap set — the entries a new agent should start from, so it inherits the useful parts of an existing brain without loading the whole history. Pruning and consolidation of near-duplicates are handled separately by memory consolidation.

​
Memory Safety Validator

Checks a candidate memory before you commit to it: contradictions against what is already stored, temporal consistency, and confidence thresholds. Deletions are additionally checked for causal chain integrity, so you do not remove an entry that a recorded decision still depends on.
This is a tool you call, not a gate that runs on every write. amfs_write does not validate — call amfs_validate first when you want a candidate checked. Validation returns findings; it never blocks a write on your behalf.

​
Multi-Strategy Retrieval

amfs_retrieve runs several searches in parallel and fuses them with Reciprocal Rank Fusion: semantic similarity (0.4), BM25 keyword match (0.2), recency (0.2), confidence (0.2) and evidence (0.3) — the outcome record, so a validated entry outranks an untested one that reads the same, and discredited entries are held back rather than ranked. A strategy whose scores carry no signal for a given query is dropped automatically instead of adding noise. Around that core:
  • Query rewriting makes first-stage recall robust to paraphrase, so a question that shares no vocabulary with the stored fact still finds it.
  • Reranking passes the candidate set through a cross-encoder before the final order.
  • A calibrated abstain gate returns nothing when the store has no answer, rather than confidently surfacing an irrelevant fact.
  • Once a learned ranking model has been trained on your outcomes, it joins the fusion and the generic signals rebalance to make room for it.
Compare configurations on real queries from Pro Tools → Retrieval Playground and Retrieval Compare.

​
Pro MCP Tools

The evaluation and repair tools — judges, incidents, fixes, replay — are covered in Behavior.