What an agent records about its own session. Stored as the content of an
agent-note memory in the text render writes, which parse reads back,
so the fields survive sync and redaction without a column of their own.
The keys for one purpose. Until settled, files may still be plaintext
or sealed under a retired key, and their owner reseals them, then calls
SqliteStore::settle_at_rest_keys.
A live memory whose stored chunk set is missing or no longer matches its
content: the chunks to embed and the hash store_chunks must be handed
back so a set split from since-changed content is refused.
Bounds the wait for a pooled read connection on this thread until the
guard drops: a read that finds the pool empty waits at most until
deadline, then opens a connection of its own, which closes on drop.
That open is bounded by BUSY_TIMEOUT, not by the deadline, and if it
fails the read waits for the pool as an unbounded read would.
Thread-local because every store read takes a connection through
reader() and a recall’s reads all run on the one thread that holds the
guard; the guard is not Send, so it cannot be dropped on another.
A row the pass could not parse. Its text was redacted in place (a chunk
row was deleted) without syncing or re-embedding it. row is rowid N,
local_seq N for sync_log, or from_id/to_id/relation_type for
relations; it and reason are redacted, since a corrupt value or an
error message can echo a secret.
Every write (via with_txn or a direct execute) goes through writer,
the single connection WAL reserves for the one process-wide writer. Reads
go through readers: a pool of separate connections in file-backed mode,
or the same connection as the writer for in_memory, where a second
connection would just see an empty, unrelated database.
What one delete removes: a memory by id, everything a Claude Code
session produced (its transcript’s raws, its subagents’, notes filed
under it), or everything filed under a stored project path.
Where the one-time pass over stored content stands. A cursor is the last
rowid its phase finished, so a restart resumes after it. Relations have
no rowid, so that phase restarts from the top; redacted rows no longer
match, which makes the rescan cheap and the outcome the same.
One row of a snapshot, in the order the exporter writes them: tombstones
first, then live rows parents before children, then the carried log tail,
and the dedup ranges last, so they are merged only once every row is.
Cap on what one traverse call returns. The walk also stops expanding
after TRAVERSE_MAX_WALK (id, depth) pairs, because a hub entity can
connect one memory to tens of thousands and the context pack that asks
only shows a handful.
A busy writer or an I/O fault says nothing about the record, so it never
settles; anything else (a parse failure, a constraint, an unsupported
shape) fails the same way on every attempt. Parsers in this crate wrap
their failures in io::Error::other, so those kinds are content.
The Claude Code session a transcript path belongs to: its file stem, or
for a subagent transcript (<session>/subagents/<agent>.jsonl) the
directory above subagents. An agent note’s source_id is the bare
session id, which is its own stem.