V4 · Behavioral Learning

Behavioral Learning Engine.

SuperLocalMemory records local behavioral signals that influence ranking without a separate LLM call for the learning step. Three phases — baseline observation, rule-based heuristics, optional ML model activation. Inspect signals, thresholds, and rank changes; optional modes and integrations retain their own network paths.

V4 AGPL v3 Local-first signals Three-phase learning Inspectable
RETRIEVAL FOUNDATION

Three retrieval mechanisms.

Semantic, graph, and adaptive channels feed the ranking pipeline. Each channel is independently inspectable.

01 · Semantic

Semantic Search

Dense vector candidates from a local embedding model. Finds memories conceptually related to a query even when keywords do not match exactly.

02 · Graph

Graph Association

Entities — files, functions, concepts — are extracted and linked. Querying one entity can surface linked neighbours via graph traversal, not keyword matches alone.

03 · Adaptive

Adaptive Re-Ranking

Local behavioral signals — access patterns, explicit feedback, recency — adjust candidate rank without an additional LLM call for the learning step itself.

LEARNING PHASES

Three-phase adaptive learning.

Recall emits local exposure telemetry. Exposure alone is not positive feedback and does not increase truth or trust scores stored in memory.

Phase 1 · 0–19 signals

Baseline

Optional reranking and channel diagnostics can record exposure. Explicit positive, negative, or corrective feedback remains separate from stored-memory confidence and trust. No model activation at this phase.

Phase 2 · 20+ signals

Rule-Based

Heuristic boosts from learned patterns: recency weight, access frequency, and trust score. The ranking influence is deterministic and inspectable — no trained model is loaded at this phase.

Phase 3 · 200+ signals

ML Model

When the learning path has sufficient valid signals and model activation is enabled, it can train on local behavioral state. Verify activation, model state, and rank impact in the installed release.

SIGNAL ARCHITECTURE

Inspectable learning signals.

Each signal type has distinct semantics. Read the source before inferring intent from a rank change.

Co-Retrieval Records co-exposure of memories in the same result set. Not correctness feedback by itself; does not update stored-memory confidence or trust.
Confidence Lifecycle Stored assertion confidence remains distinct from access telemetry. Retrieve count does not automatically upgrade confidence.
Channel Performance Requires explicit outcomes before changing quality claims for a retrieval channel. Exposure alone is not an outcome.
Diagnostics Trace the signal source, update path, and resulting rank impact. The dashboard surfaces per-signal contributions for each recall.
INTEGRATION

Integration and activation paths.

These features require hooks, background workers, or explicit configuration. Each needs independent verification for a complete deployment review.

Hook Capture Supported tool events can be captured when hooks are installed and the client supports them. Inspect event coverage, truncation, redaction, storage cost, and omissions.
Auto-Recall Project context can be injected at session start via Claude Code hooks. Verify hook installation, event coverage, and any content redaction before enabling.
Sleep-Time Work A background worker can deduplicate, decay stale signals, and conditionally retrain. Verify the worker is running and review its output before relying on its changes.
Pattern Detection Tech preferences, temporal patterns, and interests can be mined from behavioral state. Treat derived patterns as weighted context, not verified fact.
PIPELINE

The retrieval pipeline.

Query enters, candidate producers run, re-ranking applies behavioral weights, ranked context exits. Every stage is traceable.

slm recall "token expiry"
# Five candidate producers run in parallel stage candidates semantic → 12 (cosine, local embedding) bm25 → 8 (full-text index) temporal → 4 (recency window) graph → 6 (entity links) hopfield → 3 (associative)
stage re-rank fusion score · access weight · trust score · decay ✓ top-5 returned · rank deltas inspectable in dashboard
# Inspect behavioral weights on the last recall $ slm diagnostics --last-recall JWT-refresh: rank +0.12 (access_count=8, recency_boost=0.08) ✓ signal sources visible · no black-box ranking
EXPLORE FURTHER

Research and installation.

Read the published research behind the memory architecture, or follow the installation guide to deploy V4.

View Research Architecture