shared-context.md

Shared context: the human half, and the landscape

Companion to biology.md. The insect mechanisms say how a room converges; they are silent on what a trace should contain and how a reader knows it was understood. That is the human literature. The second half surveys what open-source multi-agent and multi-writer systems actually put in their schemas.

Transactive memory

Daniel Wegner, "Transactive memory: a contemporary analysis of the group mind", in Mullen & Goethals (eds.), Theories of Group Behavior, Springer (1986), 185–208; and "A computer network model of human transactive memory", Social Cognition 13(3):319–339 (1995), where Wegner himself reaches for the directory-sharing metaphor.

A transactive memory system is the individual stores plus the processes that link them, and it has three operations:

• Directory updating — maintaining a shared who knows what map. This is the actual shared object. The group's memory is the directory, not the contents.
• Information allocation — routing a new item to whoever the directory says owns that domain. Note that the encoder may then not retain the item at all.
• Retrieval coordination — knowing where to look, and in what order.

Lewis, J. Applied Psychology 88:587–604 (2003), validates the three-factor scale: specialization, credibility, coordination. Credibility is the load-bearing one here — a directory entry is useless unless the retriever trusts the owner enough not to re-derive the answer. Liang, Moreland & Argote, PSPB 21:384–393 (1995), find that training teams together improves performance and that the effect is mediated by directory quality. Ren & Argote, Academy of Management Annals 5:189–229 (2011), is the review.

The failure modes are named: directory staleness, shared-information bias (Stasser & Titus again — the group discusses what everyone already knows), and the Google effect (Sparrow, Liu & Wegner, Science 333:776–778, 2011 — people who expect an external store to persist remember where rather than what, which is the intended behaviour until the store is lossy).

What this workspace would have to represent. Roster answers who is here. Nothing answers who knows what. The hidden-profile failure recorded in the live experiment is precisely a transactive-memory failure: scout's refuting fact was in the room, at a citable sequence, and nothing could route the floor to the member holding it. AgentThreshold.affinity is the shape of a directory entry — an agent, a topic, a weight — supplied statically by the host rather than folded from the transcript. Folding it is a pure operation over traces the crate already reads, and specialization and credibility both have obvious estimators in it: who deposited grounded evidence on a topic, and whether their deposits were later cited or refuted.

Distributed cognition

Edwin Hutchins, Cognition in the Wild, MIT Press (1995); "How a cockpit remembers its speeds", Cognitive Science 19(3):265–288 (1995).

The formulation: cognition is "computation realized through the creation, transformation and propagation of representational states across representational media". The unit of analysis is the system — people plus instruments plus procedures plus layout — not the individual.

In the ship's-bridge case a position fix is computed by a chain in which nobody holds the whole computation: pelorus operators sight landmarks and report bearings by voice, a recorder writes them into a durable timestamped bearing record log, a plotter transfers them to the chart with a protractor whose mechanical constraint performs part of the trigonometry. The chart is not a display of the answer; it is where the computation happens. Hutchins's decisive move is that the system's cognitive properties follow from the physical properties of the media — a written log is reviewable, persistent and inspectable by a bystander; a spoken bearing exists only at the moment of utterance. When the gyrocompass failed, the crew reorganized the propagation pathways, not anyone's knowledge.

The cockpit paper makes the memory point sharpest: V-speeds are remembered by the cockpit as a system, partly in heads, partly in a speed-card booklet, and critically in the speed bugs physically set on the airspeed dial. The bug converts a remembered number into a perceptual judgement — is the needle past the bug? — and that is the deepest available lesson for a transcript: the best shared memory changes the task from recall to perception.

What this workspace would have to represent. The attributed transcript is a bearing log and it is already the right shape. What it lacks is the second medium. Every intermediate representation currently has exactly one form: prose in a SessionMessage, re-parsed on every read. A folded standing — which topic carries, who supports it, how far it is from quorum — is computed and thrown away each step. The bench harness already found this: the fix for models coining #rollout and #rollout-strategy was to show the room its own standings in the prompt. That is a speed bug, and it lives in an example rather than in the library.

Grounding

Herbert Clark & Susan Brennan, "Grounding in communication", in Resnick, Levine & Teasley (eds.), Perspectives on Socially Shared Cognition, APA (1991), 127–149.

Common ground is what participants mutually believe they share, and each contribution has a presentation phase and an acceptance phase in which the addressee supplies evidence of understanding. The grounding criterion is that both parties mutually believe understanding was reached "to a criterion sufficient for current purposes" — task-relative and negotiable, not absolute, which is a far cheaper contract than "shared state must be identical" and the right one for agents too. The principle of least collaborative effort says participants minimize joint effort, not their own.

The eight media constraints are the part that reads as a schema checklist: copresence, visibility, audibility, cotemporality, simultaneity, sequentiality, reviewability, revisability. A shared append-only transcript grants reviewability — the whole point — and simultaneity, and denies copresence, visibility and audibility, which means every cheap back-channel signal of uptake is unavailable and acknowledgement must be made explicit and paid for. It grants sequentiality only where a total order is enforced. Revisability is a genuine hazard rather than a feature: if a message can change after it is read, an agent citing it is citing something that no longer says what it said.

What this workspace would have to represent. Sequentiality is bought already — Sequence is a total order and same_conversation scopes it — and revisability is denied by construction, which is the correct choice and worth stating as one. Acknowledgement is the gap: repetition_cap counts how many "distinct peers have acknowledged a point", but acknowledgement is inferred from a trace being deposited, never stated. Trace.cites is the nearest thing to positive evidence of uptake and is used for grounding rather than for grounding-in-the-Clark-sense. A trace kind meaning "I have read this and it changes nothing" is cheap, and is the only way a room can distinguish silence from assent.

Boundary objects and awareness

Boundary objects. Star & Griesemer, Social Studies of Science 19(3):387–420 (1989): objects "plastic enough to adapt to local needs … yet robust enough to maintain a common identity across sites", weakly structured in common use and strongly structured in local use. The point is coordination without consensus. Four types: repositories, ideal types (useful because imprecise), coincident boundaries, and standardized forms — the last being the one most prone to becoming a rigid, information-losing artifact. Star's own corrective, Science, Technology & Human Values 35(5):601–617 (2010), insists the interpretive flexibility is not optional.

Workspace awareness. Gutwin & Greenberg, CSCW 11(3–4):411–446 (2002). Awareness is a two-tense structure — who/what/where now, plus how/when/who-did- that historically — and it is gathered three ways: consequential communication (information given off unintentionally by the act of working), feedthrough (observing the effects of someone's actions on shared artifacts), and intentional communication (explicit telling, the expensive fallback). Feedthrough is sematectonic stigmergy under a different name, and the design rule is to maximize the first two so as to spend as little as possible on the third.

Situation awareness. Endsley, Human Factors 37(1):32–64 (1995): perception, comprehension, projection. A raw event feed delivers only perception. Team situation awareness is explicitly not the union of individual awareness — over-broadcasting the non-shared parts degrades it.

What this workspace would have to represent. TeamBriefing and SessionContext are boundary objects and are correctly kept separate from history. ThreadLine is the beginning of a repository index. What does not exist is feedthrough: an agent's tool calls and artifact edits are invisible unless it narrates them, so a room's only awareness channel is the expensive one. That is the host's to emit, but the record kind it would emit into is this library's to define — and today every record is a chat message.

Global workspace

Baars, A Cognitive Theory of Consciousness, CUP (1988); Mashour, Roelfsema, Changeux & Dehaene, Neuron 105(5):776–798 (2020). Massively parallel specialized processors compete for a capacity-limited workspace; the winner's content is broadcast back to all of them. Ignition is the nonlinearity: "the sudden, coherent, and exclusive activation of a subset of workspace neurons … with the remainder inhibited" — all-or-none, self-sustaining after the stimulus ends, and exclusive by lateral inhibition.

This is worth recording because the architecture is already that shape. bids is the competition, floor_holder is the argmax, and appending to the shared transcript is the broadcast. The one-message-one-turn rule is not a concession to safety with a performance cost; it is the same design a brain uses, and the lateral inhibition is !object. Self-sustaining ignition is the one part missing: a decision that carries has no minimum dwell, so a late trace can move standings the step after they settled — which is what commit_boundary and the one-way Deliberate → Commit transition exist to bound.

The open-source landscape

Read for schema, not for marketing. Full notes with URLs are in the pull request that added this file; what follows is the transferable part.

System	The lesson
Zulip	Conversation scope is a mandatory, total, mutable string on every record, addressed as (channel, topic). Mutability is the feature: agents mislabel, and re-filing is cheaper than fragmentation. Per-participant state is one row per (reader, message) with flags, so per-topic unread counts fall out and no threaded-receipt mechanism was ever needed.
Slack	The counterexample. Threading is one nullable thread_parent_id, added years after channels, optional and secondary — which produced non-uniform adoption, a second unread model, and finally "also send to channel" as an escape hatch. Optional structure is worse than none.
Matrix	prev_events (what the sender had seen) and auth_events (what makes the write legitimate) are different edge sets over the same node. Event id is a content hash, so dedup, integrity and citability are one field. Threads via m.thread are server-side aggregatable, so a client gets {latest, count, participated} without downloading history. And MSC3771 had to make read receipts per-thread, because "read receipts and read markers assume a single chronological timeline".
Kafka	Immutable log plus per-reader offset is the entire architecture of several readers at different positions. Compaction is offset-stable — retained records keep their original offsets, so compaction leaves gaps and never renumbers, and every held cursor and citation stays valid. Tombstone lifetime is bounded by the slowest reader, not by a fixed TTL.
MetaGPT	A shared message pool with two orthogonal addressing axes: send_to (who it is for) and cause_by (what it is about), with roles subscribing on the second. Also instruct_content: a structured, machine-checkable payload carried beside the prose.
AutoGen / AG2	ChatMessage versus AgentEvent — two tiers over one stream, separating what enters another agent's prompt from what is only telemetry. OnContextCondition gives a deterministic routing path beside the LLM one, so selection does not always cost a model call.
LangGraph	A reducer per channel: different fields want different merge policies in one session. add_messages is an ID-keyed upsert, so edit and redact are ordinary appends. Long-term memory is namespaced by a hierarchical tuple with prefix search.
Letta / MemGPT	Memory blocks can be attached to several agents at once — genuine shared mutable state — and the docs rank the mutation primitives by concurrency safety: append is safe, replace is a compare-and-swap that fails loudly, wholesale rewrite loses updates and needs an owner. Say which is which in the API.
Zep / Graphiti	Bi-temporality: created_at/expired_at (when the system learned it) is separate from valid_at/invalid_at (when it was true). Nothing is deleted; a contradiction supersedes. Point-in-time replay needs both axes.
A2A	Message (conversational, ordered) versus Artifact (durable, named output) — two record kinds in one log. Conflating them forces readers to re-derive results by scanning prose.
MCP	Update notifications carry only a URI, never a payload: invalidate, do not push, so there is no ordering or merge problem. annotations.audience and annotations.priority are a ready-made vocabulary for per-reader projection and context-budget eviction.
Claude Code / OpenCode transcripts	The session is a tree, not a list: parentUuid linkage makes a rewind a branch, and a summary record carries leafUuid pointing at the tip it replaces. Compaction is an added node, never a deletion. isSidechain keeps subagent traffic out of the derived view while staying in the same store.
Blackboard / HEARSAY-II	Levels of abstraction as first-class partitions of the shared board, hypotheses carrying credibility ratings integrated across levels, and knowledge sources triggered by changes to the board rather than by a pipeline. The classic critique is that the control problem then dominates — which is what "auto speaker selection costs a full transcript read per turn" is a restatement of.
Linda tuple spaces	rd (observe) and in (claim, destructively, with a uniqueness guarantee) are different operations. Generative communication decouples producer and consumer in time, space and identity — nobody addresses anybody, which is stigmergy from the coordination-language side.
Mem0	The negative result. An LLM arbitrates ADD/UPDATE/DELETE over shared state, which makes the write path non-deterministic and unauditable. Keep the model in the content and deterministic code in the merge.
Cognition versus Anthropic	Cognition: share full traces, because a summarized message loses the reasoning and parallel writers then make conflicting implicit decisions. Anthropic: repeated condensation plays telephone with citations, and token usage explains most of the performance variance. Both are right, and the only shape that satisfies both is one durable full-fidelity log plus cheap per-reader projections — which is already this library's shape and is the strongest thing to say about it.

What the landscape says this workspace already gets right

Worth recording, because it is the argument for the design rather than against it.

• One log, many projections. Every framework above either floods context (broadcast to all) or fragments it (pairwise isolation, per-agent retrieval). project_session, project_for and prepare_delta narrow presentation without narrowing availability, which is the shape Cognition and Anthropic jointly argue for and neither implements.
• Deterministic merge. Every fold here is pure and replayable. Mem0's LLM-arbitrated writes are the counterexample, and CrewAI's implicit retrieval assembly means you cannot reproduce what an agent saw.
• Attribution is never collapsed. SessionAuthor survives projection. This is P4, and it is the one thing every chat-shaped framework gets right and every memory-shaped one loses.
• No second journal. The charter's refusal is the same discipline as Kafka's single log and event sourcing's single source of truth.

What it says is missing

In the order the work plan takes them:

1. Evidence cannot argue against a topic, only against an advocate — the α(v) gap from biology.md, and ADR 0003.
2. Grounds are counted, not weighed, so a citation of a citation of an opinion is worth a citation of a fact — the cascade condition, and ADR 0004.
3. No who-knows-what directory, so a room cannot route to the member holding the fact.
4. SessionMessage drops parent and carries no structured payload, so thread narrowing cannot be a caller's fold and traces are re-parsed on every read.
5. One scalar watermark per agent, which is only correct on a single timeline — the retrofit Matrix had to do.
6. No compaction that keeps citations resolvable, and no supersession.