Multiplayer agent harness for work.
Memory records and audience isolation
One notebook representation
Built-in memory uses versioned records: stable ID, text, sensitivity, source
scope/session references and an incomplete-provenance flag (sourceUnknown).
Markdown is the rendered editing/display format, not a second source of truth.
Headings and freeform prose are records too.
Postgres stores the authoritative snapshot in memory_revisions.records. New
revisions derive body from those records for compatibility. A legacy revision
with NULL records imports its body as unknown-source records; its sources are never
guessed from where the notebook happens to live. An older writer during a rolling
deployment can still produce a NULL snapshot, which takes that conservative path.
The additive migration preserves existing revisions.
The workspace backend stores one JSON snapshot at its existing reserved memory path. Existing Markdown is imported on read and migrated on the next write. The memory API still returns Markdown. Malformed structured data fails closed rather than silently falling back to potentially stale text. Old file-only binaries cannot read the new snapshot format; export through the memory API before downgrading one.
Capture and curation
Automatic extraction classifies each batch as ordinary, unknown, sensitive
or restricted. Explicit capture uses the same labels. Missing or malformed
classification and model failure produce unknown. Classification is a model
estimate, not proof of harmlessness.
Runtime context supplies source scope/session, never model-authored source claims. Captures inherit known input restrictions. Untracked history, tools, attachments or incomplete dependencies add unknown provenance. Cross-scope personal copies keep their original source. Duplicate capture can tighten metadata without adding text.
Built-in consolidation edits records by stable identity under a revision check. Each model request sees only facts with identical sensitivity and provenance; updates and additions inherit that group's metadata. Unchanged records keep their IDs and metadata. Separate-source facts never share a consolidation prompt. Bookkeeping is a snapshot counter, not a Markdown marker that rewrites every fact. A concurrent edit wins over a stale consolidation. Opaque providers retain control of their own consolidation; generic whole-Markdown automatic rewriting is removed.
An arbitrary whole-notebook rewrite cannot establish which facts its new text came from. It conservatively inherits the old notebook's combined restrictions. Restoration likewise inherits the restrictions of all intervening revisions. Neither operation declassifies information. Clearing removes current content, not historical revisions.
Fresh reads
One disclosure view serves recall snapshots, search, primitive tools, memory-file redirects and agent memory APIs. Provider routes apply it per provider. Structured search runs over filtered records, not an unfiltered provider query.
The actor must retain access to every recorded source. Carrying memory into another conversation requires current Open policies at both ends and a known destination audience. An ordinary personal fact can follow its owner into an eligible internal Open room; sensitive or restricted facts require every recipient to be entitled to the sources. Ordinary does not mean public. Unknown provenance stays within the notebook's existing narrow audience. Lookups fail closed, and authorization results are not cached between reads.
A partially visible notebook cannot be rewritten. Structured edits require a revision check. Historical reads apply a conservative restriction floor from newer revisions; restore checks the complete intervening history in storage.
Opaque external providers and the optional scratch-promotion workflow remain compatibility paths, not sources of trusted cross-context provenance. Their unclassified content stays narrow. Scratch captures with foreign dependencies and cross-origin personal copying are refused. Scratch bookkeeping lives in snapshot metadata, not fact text; a no-change promotion preserves every fact's provenance. Direct MCP tools and authorized private administrator inspection are separate access paths.
Retained working context is separate
Fresh-read restrictions do not erase information legitimately learned by the same audience. Source revocation, notebook edits or disabling recall do not reset that audience's conversation.
A durable session checkpoint records the destination/audience identity hash and a replay cutoff. An actual audience change advances the cutoff: earlier model history, summaries, native harness state, retries, approvals and historical tape cannot re-enter the next turn. Later turns retain post-cutoff context normally. An absent checkpoint establishes a baseline rather than deleting old context. Project conversations keep their existing member-inherits-history behavior.
Agent conversation APIs and subagent reads also respect the cutoff, including derived titles and metadata. Human transcripts and administrator audit records remain intact. Unversioned external prior turns cannot safely be imported after a cutoff. This is next-turn isolation, not cancellation of a model request already in flight, nor retroactive deletion of exported or copied artifacts.
Verification limits
Synthetic tests cover storage migration, metadata-only capture, consolidation, revision races, restoration, provider routing, fresh authorization and durable context boundaries. Live HTTP tests use synthetic facts through the real server and Postgres. Deterministic model responses isolate consolidation behavior; these tests do not establish classifier accuracy or qualify every real model and Slack configuration.