This page describes the active-memory components introduced in V3.1. The current release includes local feedback, outcome, co-retrieval, and adaptive ranking machinery. It does not guarantee that ranking improves for every corpus, and it does not treat result exposure as proof of relevance.

Recall can emit local telemetry that later feedback or outcomes can qualify:

Recall → Exposure telemetry → Explicit feedback/outcome → Adaptive ranking

The feedback/ranking path itself can operate locally. Provider-backed extraction or other optional integrations have separate network and token behavior.

Signal counts and training readiness depend on actual use and explicit outcome coverage. V3.7 publishes no time-to-training guarantee.

When memories are retrieved together repeatedly, they form implicit connections. The system learns that these memories are related — even if they don't share keywords.

Recency, access, trust, and lifecycle state can affect internal ranking utility. They do not change relevance_score or memory_confidence into answer probabilities. See Score Contract v2.

Current candidate producers are dense semantic, BM25 lexical, temporal, Hopfield associative, and spreading activation. Entity-graph information is a post-fusion score enhancement in the current implementation.

When new content arrives, the system measures how "surprising" it is relative to existing memories. High-entropy content (genuinely new information) gets prioritized for deeper indexing.

SLM can automatically detect and store decisions, bug fixes, and preferences from your conversations:

slm observe "We decided to use PostgreSQL because of JSONB support"
# Auto-captured: decision (confidence: 0.75)

Detection patterns:

• Decisions: "decided", "chose", "switched to", "using X because"
• Bug fixes: "fixed", "root cause was", "resolved by"
• Preferences: "always use", "prefer", "convention is"

At the start of every session, the system can automatically inject relevant context:

slm session-context  # Returns top-10 relevant memories for current project

slm hooks install  # One-time setup

This explicitly installs supported Claude Code hooks. The current package installers do not install hooks or edit IDE configuration implicitly.

Three new MCP tools for AI assistants:

A background maintenance process runs periodically:

• Decays confidence on unused memories
• Deduplicates near-identical facts
• Generates behavioral patterns from accumulated data
• Auto-retrains the ML ranker when enough signals accumulate

Trigger manually: slm consolidate or via the dashboard's Learning tab.

The Learning tab shows:

• Signal count and phase progression (0 → 20 → 200)
• Tech preferences learned from your memories
• Temporal patterns (when you work on what)
• Channel performance (which retrieval channel works best)

The Behavioral tab shows:

• Learned patterns with confidence scores
• Outcome tracking (success/failure/partial)
• Cross-project pattern transfer

Adaptive ranking needs representative feedback and held-out evaluation. A retrieved fact is only an exposure event until explicit feedback or a qualified outcome exists. V3.7 does not publish a uniqueness claim, guaranteed learning curve, or calibrated answer-confidence result for this subsystem.

Part of Qualixar | Author: Varun Pratap Bhardwaj