- Assertion
- One structured fact extracted from a source, written as subject → predicate → object (for example “Atlas archive → opens at → 09:00”). Every assertion points back to the text that supports it.
- Canonical record
- The authoritative copy of a fact, kept in SQLite. Search data in LanceDB and Kùzu is derived from it and can be rebuilt from it.
- Projection
- A copy of canonical data reshaped for one kind of search: vectors in LanceDB for meaning, a graph in Kùzu for relationships. Projections follow SQLite; they never decide anything.
- Mutation
- Any change to memory: inserting a source, correcting it, rolling it back or purging it. Each mutation has an ID and a status you can poll.
- Admission
- The moment MESA accepts a mutation and records it durably. The API answers 202 at admission; processing happens afterwards.
- COMMITTED
- The successful final state of a mutation: SQLite, LanceDB and Kùzu have all been written. Only then is the new memory fully searchable.
- DLQ (dead-letter queue)
- Where work goes after its retries run out. It stays there, visible, until an operator replays it. Nothing in the DLQ counts as stored memory.
- Principal
- The caller identity behind an API key. Roles are granted to a principal, never to a request.
- Tenant · workspace · dataset
- The three nested scopes data lives in. A tenant contains workspaces, a workspace contains datasets. Permissions are granted at any level and inherit downward.
- ACL and roles
- Who may do what. Roles are OWNER, WRITER and READER; purge and rollback need their own explicit permission on the dataset.
- Session
- Created by the server for one principal and one agent. It fixes which datasets can be read for its whole life; the client cannot widen it.
- Agent
- The AI persona or worker a session acts for. It labels context; it is not a security boundary.
- Revision and chunk
- A document is stored as immutable revisions; each revision is split into exact text chunks. A search result names the document, revision and chunk it came from.
- Embedding
- A list of numbers that represents the meaning of a text, produced by an embedding model. Query and stored embeddings must come from the same model and size.
- BM25
- A classic keyword ranking method. It rewards passages that contain the query’s words, especially rare ones. MESA runs it on SQLite FTS5.
- RRF (reciprocal rank fusion)
- A way to merge several ranked lists: each list gives a candidate 1 / (60 + its rank), and the sums are compared. It uses positions, not raw scores, and is not a probability.
- Validation mode (MESA_TIER3_MODE)
- How strictly extracted facts are checked before they become memory: 0 = rule-based checks, 1 = plus one model validator, 2 = two different models must agree.
- V3 and V4
- Two API generations in the same package. V4 is the current design with datasets, sessions and the ledger; V3 is kept only for compatibility.
- MCP (Model Context Protocol)
- An open protocol that lets AI tools call external services. MESA offers an MCP gateway so an agent can search and write memory within its session scope.
- Profile B
- The end-to-end certification scenario that moves frozen legal data from MESA Data into Core and checks every receipt. It has not passed yet.
- NO-GO (production status)
- The project’s own verdict that Core is not yet approved for production use. Remaining gates are listed on the status page.