Method
Method
llearn's working hypothesis is that learning in new domains can be accelerated when familiar patterns from known domains are recognised, trusted, and transferred into an agentic harness. The method is the operating layer beneath the visible tools: it turns notes into context, context into routes, and routes into a cheaper, faster, more accurate next session.
The loop
01
Capture
Work starts with ordinary material: notes, documents, voice capture, code, source fragments, and whatever format the operator naturally uses.
llearn wraps each source item as an artifact so provenance stays attached before any synthesis happens.
02
Ingest
Artifacts are broken into atomic nodes: distinct claims, decisions, observations, code fragments, or pieces of prose.
Each node is embedded, vectorised, and linked to nearby nodes with typed relationships. One node might enhance, contradict, update, or operationalise another. Cosine similarity provides the first familiar-pattern signal. It is not the final source of truth.
03
Traverse
Agent interactions are required to draw from the substrate, not loose prompt memory.
Traversal follows typed links through the graph toward stable core anchors: high-trust principles that ground the answer before the system returns to the concrete task. The path and interaction type are logged so the system can learn from use.
04
Curate
Nodes and links carry weights from 0.001 to 1.0.
Curating workers review interaction logs and adjust those weights against an explicit weighting table. Knowledge and operational signals stay separated: knowledge weights reflect what the operator is likely to trust and understand; operational weights reflect the complementary behaviour the harness should provide.
Current stage
The high-level design has made the approach technically plausible. The work is now in evals: deconstructing strong agent harnesses, reconstructing the best parts, and setting a baseline for cross-domain accuracy, human-in-the-loop confidence, and operator understanding before the full system is built.