Retrieval Maturity Levels

A ladder for how a knowledge store retrieves, framed as five levels = five questions, each answering a distinct retrieval need:

Route — "Find it by an exact word or name?" Files + folders + a Context Routing file; native keyword/filename search. Most people never need to leave here.
Topic wiki — "Pull everything on a topic together?" Ingested notes cross-linked into indexed wiki pages; retrieval is "follow the trail and read the page."
Semantic — "I searched different words than I wrote?" Semantic Retrieval by embedding/meaning instead of exact match.
Knowledge graph — "Trace relationship chains across a cast (a CRM-like web)?" Knowledge Graph Retrieval with typed entities and relationships.
Autonomous — "Consolidate on its own while I'm away?" An always-on "Brain OS" that syncs and refreshes memory without you (e.g. Gbrain).

The load-bearing claim is not that higher is better. Complexity climbs with the level; capability does not automatically. The correct move is the lowest level that removes a pain you actually have — "most people land at 1–3," and moving up is only "better" when it fixes a felt pain point. This is Agentic Simplicity applied to retrieval architecture: no pain, no climb. A single store is also heterogeneous — one folder can be level 2 while another is level 4; you pick per folder based on the data and how it's queried (Query-Shaped Storage).

Claims

Each level answers a different retrieval question — pick by the question you actually ask. principle — durable: the level is a function of the query shape, not a status tier to ascend.
Complexity climbs as you go up the levels; capability does not automatically follow. principle — durable: this is the whole corrective — the flashy level is usually not the needed one.
Choose the lowest level that fixes a pain you have; no pain, no climb. best practice — context: operators tempted by knowledge-graph demos; "best" is the least complex retrieval that removes a felt pain, so absent pain the best choice is to not climb. Most people land at 1–3.
Moving up a level is not inherently better — moving down isn't either. observation — a factual reframing: it's about matching the pain point, not ranking the levels.
A single store can mix levels folder by folder. best practice — context: real brains hold data of different shapes; "best" is to match each folder's retrieval style to its data and queries rather than forcing one style everywhere.
Level 5 (always-on autonomous ingest) risks too much context doing more harm than good — manual control over what's ingested is a defensible default. best practice — context: when consolidation is automated, the operator loses the Evergreen vs Volatile Context gate; staying manual is "best" precisely when signal/noise control matters more than hands-off convenience.
Level 3 (semantic) is not a terminus — semantic search fails on exact identifiers, so the practical default is Hybrid Retrieval (keyword + semantic fused). best practice — context: production memory whose queries mix meaning with precise strings; a second, independent source (Damian Galarza, on OpenClaw) sharpens the ladder by showing level 3 alone breaks, and that even the keyword half can win outright — the Claude Code team dropped a vector DB for grep + agentic search because it was better and easier to maintain.

AI Second Brain — the store this ladder describes.
Context Routing — level 1, and the substrate every higher level still sits on.
Semantic Retrieval — level 3.
Hybrid Retrieval — the refinement of level 3: keyword + semantic fused, because semantic alone breaks on exact identifiers.
Knowledge Graph Retrieval — level 4.
Query-Shaped Storage — why the level is chosen per-folder by query shape.
Compiled Knowledge Base — the Level-2 "topic wiki" seen as an ingestion philosophy: compile-once-and-integrate vs retrieve-and-re-synthesize per query.
Agentic Simplicity — "no pain, no climb" is "add complexity only when it demonstrably improves outcomes," restated for retrieval.
Distillate: Every Level of a Claude Second Brain Explained
Distillate: Fable 5 + Karpathy's LLM Wiki is Basically Cheating — a hands-on Level-2 walkthrough (the Karpathy "LLM wiki").
Distillate: How AI Agents Search Their Memory — Hybrid Retrieval, in Practice (OpenClaw) — level-3 refinement + the Claude-Code-drops- vectors corroboration.

Retrieval Maturity Levels

Claims

Related

Linked from