Glossary
The words, before the paper uses them
Short definitions of the terms on this site. Each one points at the page that shows the thing rather than only naming it.
Engram
The unit FluctlightDB stores. An engram is a piece of content together with its context, its salience and its provenance, and edges to memories that were written beside it. It is not a row in a fixed schema and it is not a bare vector.
experience()
The write. An agent records what happened — a tool result, a user fact, an observation — at the moment it happens, with context, salience and provenance. Nothing is extracted from a transcript afterwards.
Cue recall
activate(cue) asks what this cue should bring back, not what is nearest. The question is the cue. The answer is a ranked list of memories. recall() is a different call with a different return shape; activate() is the documented path.
Spreading activation
Inside one activate() call, full-text search, vector neighbours and a walk of the memory graph are fused. Memories that were linked when they were written can come back together even when a single channel would have missed one of them.
Provenance
Where a memory came from. A memory written with verified=True and a source_uri is something the agent can point at. A memory without that is something it was told. At recall, a verified source ranks above hearsay.
Salience
A weight, passed to experience(), for how much a memory matters. It is recorded with the memory. It is not a score the caller computes again at query time.
Consolidation
Replaying and compacting what has been stored. consolidate() runs one pass. sleep() replays, strengthens and forgets. connect_agent() can do this on its own. checkpoint() is a different call: it makes the current generation durable. The neuroscience-named modules around this machinery are experimental and are not a promise of new neuroscience.
checkpoint()
The durability call. It writes a new generation and advances the pointer to it. Writes also go to a write-ahead log, and a process killed mid-write leaves the previous generation intact. Nothing is durable until checkpoint() returns.
Brain directory
One directory per agent. connect_embedded(path) opens it, creating it if needed, and the engine runs inside the Python process. There is no server to start for that path. Several agents can share a directory, and they should not: provenance separation across agents in one brain is poor, and the project says so.
Lexical recall
Offline, with no embedder of your own, a cue has to share tokens with what was stored. Paraphrase recall needs a semantic vector you pass on both the write and the cue. No embedding model ships with the engine.