MESA LEARN / A CONNECTED COURSE

Understand memory.
Build with evidence.

From your first question about AI memory to a search system you can run and inspect. One course, concrete examples, no assumed background.

14 lessons · read in order or explore by question

HOW TO USE THIS COURSE

Learn one idea.
Then make it work.

  1. 01 / Follow the example

    The fictional Atlas archive connects the lessons. Track the same source as it becomes a record, a search hit and evidence.

  2. 02 / Check your reasoning

    Calculate the worked examples, then try each question before opening its explanation.

  3. 03 / Build and verify

    Finish with a Python/SQLite lab. Keep the source IDs, inspect failures and measure what actually changed.

THE COURSE LIBRARY

Start at zero.
Or follow your question.

Markdown index ↗
01

Start from zero

Context, RAG and source-linked records

  1. 01

    AI memory: what survives after a conversation?

    Distinguish model weights, the current context and persistent records. Design a memory write before choosing a database.

    memorycontext windowsessionkalıcı hafıza
    5 min
  2. 02

    RAG and memory: build the request step by step

    Trace a source from retrieval into an answer. Learn why RAG and memory overlap, how context budgets work, and where grounding can fail.

    RAGchunkingcontext budgetparçalama
    5 min
  3. 03

    From documents to claims: preserve what extraction loses

    Represent exact text, entities and assertions without confusing an extracted claim with verified truth. Work through a source-linked record.

    assertionentityschemachunk
    5 min
02

Learn to retrieve

Represent, rank, fuse and follow relations

  1. 04

    Embeddings and semantic search, with numbers

    Calculate similarity, distinguish dense and sparse representations, and understand why nearest neighbors are neither facts nor access decisions.

    embeddingdensesparsecosine
    5 min
  2. 05

    BM25: why a rare word can beat repeated words

    Work through term frequency, rarity and length normalization. Tune tokenization and field weights before chasing a larger model.

    BM25IDFtokenizationFTS5
    5 min
  3. 06

    Hybrid retrieval: candidates, fusion and reranking

    Design a search pipeline where lexical and semantic lanes contribute different evidence. Diagnose whether recall, fusion or reranking is failing.

    hybridrerankingcandidate recallcross encoder
    5 min
  4. 07

    Reciprocal Rank Fusion: calculate and inspect every vote

    Fuse two ranked lists by hand, handle duplicates and ties, and understand why an RRF score is not confidence.

    RRFrank fusionweightedk=60
    5 min
  5. 08

    Knowledge graphs: follow relations without inventing proof

    Build a source-linked graph, trace a two-hop question, and distinguish graph traversal from GraphRAG summaries and valid inference.

    graphGraphRAGmulti hopKuzu
    5 min
03

Build trustworthy systems

Evidence, permissions, durability and change

  1. 09

    Evidence and provenance: what a citation actually proves

    Build a verifiable citation with source revision and exact span. Distinguish integrity, provenance, claim support and source truth.

    provenancecitationhashSHA256
    5 min
  2. 10

    Shared memory: who may retrieve what?

    Trace identity, roles and dataset scope through retrieval, caches and graph paths. Test isolation before measuring relevance.

    authorizationRBACtenantscope
    5 min
  3. 11

    Canonical storage and projections: survive a partial write

    Trace one mutation through SQL, vector and graph stores. Understand durable admission, retries, idempotency and projection rebuilds.

    SQLiteoutboxidempotencyprojection
    5 min
  4. 12

    Changing facts: validity, contradiction and forgetting

    Distinguish when a statement applies from when it was recorded. Handle corrections and conflicts without equating age with invalidity.

    temporalvaliditydecaycontradiction
    5 min
04

Measure and build

Judge retrieval and run your first lab

  1. 13

    Evaluate retrieval: recall, MRR, nDCG and latency

    Calculate ranking metrics on a judged list, separate quality from speed, and design an experiment that does not grade its own training examples.

    evaluationMRRnDCGrecall
    5 min
  2. 14

    Build a small search lab, without an LLM

    Run a complete Python/SQLite example, predict its results and add a judged regression case. Connect the course concepts to an inspectable program.

    PythonSQLiteFTS5lab
    5 min

AFTER THE COURSE

Take the ideas into a real system.

These lessons explain general concepts. For MESA’s implemented behavior, use the version-pinned API, architecture and startup guides.

Open the MESA quickstart →