Learnings

A learning is a candidate piece of per-agent memory that has been surfaced for human review and possible promotion to the institutional zone of a datapack. Learnings are the bridge between what an individual agent has learned in its own conversations and what every agent in the datapack can read.

The promotion flow

Meko separates what an agent learns from itself from what the wider system trusts. Memory starts private and per-agent. Promoted content becomes available to every agent in the datapack.

The pipeline has three stages.

1. Extraction

As an agent has conversations, Meko extracts entities, relationships, and facts and stores them as memories scoped to that agent. The agent owns this memory; no other agent in the datapack can read it.

2. Learning identification

A subset of memories is flagged as a learning, meaning it is a candidate for sharing across agents. Today this identification is a manual step performed in the Meko UI; the Learnings tab surfaces candidates and shows whether each candidate has already been promoted. Future releases will introduce automated identification heuristics.

3. Promotion

A reviewer promotes a learning into one or both of the institutional destinations:

• Collective Memory — entity and relationship content that joins the shared graph and vector store. Other agents retrieve it the same way they retrieve their own memory: through semantic and graph search.
• Shared Knowledge — document-style or factual content that joins the knowledge base and is retrieved through RAG.

Both destinations sit inside the same datapack, so promoted content is bounded by the datapack's tenancy. A learning is never shared across datapacks.

Per-agent versus institutional

This separation is what makes Meko's collective memory model auditable. Every promotion is an explicit decision by a reviewer, captured in the trace for that datapack.

Why two institutional destinations

Collective Memory and Shared Knowledge serve different retrieval patterns:

• A learning like "Customers in the Northeast region prefer Tuesday delivery windows" is best stored as Collective Memory, because the retrieval value comes from graph traversal and entity similarity.
• A learning like "The 2026 Q1 incident response playbook supersedes the 2025 version" is best stored as Shared Knowledge, because the retrieval value comes from chunked semantic search over a document corpus.

Reviewers choose the destination at promotion time. A learning can be promoted to both.

Next steps

• Memory — how per-agent memory is extracted and stored
• Knowledge bases — how Shared Knowledge is indexed and queried
• Work with memory — the hands-on guide that walks through extraction, identification, and promotion