decision

ADR 0008: `memory` goes `established` on two-vendor convergence - but the evidence bar does not move

Context#

ADR-0007 created brain/topics/memory.md at Status emerging on one source (S6, OpenAI's Dreaming: Better memory for a more helpful ChatGPT), and named its own revisit trigger: "the parked Anthropic source is distilled (which should take memory to established)".

That source is now ingested as S7 (Memory and dreaming for self learning agents, Anthropic, 2026-05-21): 23 gated nodes, 19 corroborated, slides plus a live product demo.

AGENTS.md defines established as "two or more corroborating sources". On a literal reading the promotion is automatic. It is worth an ADR anyway, because the obvious reading of "corroborating" is wrong here in a way that would quietly inflate confidence across the whole topic.

The two sources agree to an unusual degree:

• Both ship a background batch process that curates memory a session wrote.
• Both decouple it from the session and run it on its own clock.
• Both independently named it dreaming.
• Both are T2 vendors describing their own products.

Decision#

Promote memory to established, and record explicitly that the promotion is about the design and not about the results.

The topic note, claims.md claim 58 and the INDEX row all carry the same qualification: two independent vendors converging on one architecture is strong evidence that this is the natural answer to maintaining memory across sessions. It is not evidence that the architecture works.

The independence rule is satisfied, and it buys less than it usually does. AGENTS.md requires external evidence to be independent of the original source - not the same author, organisation or commercial interest. S6 and S7 clear that bar completely: different organisations, different products, different system classes (consumer chat assistant vs multi-agent platform), competing commercial interests. What they do not clear is disinterest. Two vendors agreeing about their own products are two positions that happen to rhyme, and convergence would look identical if both designs were wrong.

So the status advances on the axis the status actually measures - is this a recognisable, recurring area with more than one source? - while the confidence column on individual claims does not move. Claims 59-64 land at emerging or needs-check, not corroborated, because only claim 58 has two legs; the rest are S7 alone.

The finding that mattered more than the promotion#

S7 was expected to close the topic's headline open question, inherited from claim 24: naive episodic append-and-retrieve memory hurt agent long-horizon reliability on 6 of 10 models (T3 preprint, measured), while S6's numbers come from a chat assistant. Nobody has measured whether a maintained memory helps an agent.

S7 is the first source in this brain on the right side of both axes - maintained memory, agent platform, long-horizon, multi-agent - and it does not close the gap.

The evidence bar moved down, not up. S6's figures were recovered from the publisher's own chart specs: exact numbers, undisclosed method. S7's are marketing testimonials with no baseline, no sample size and no eval set. A second source made the topic broader and its evidence weaker, and recording that is more useful than letting "97% fewer first-pass errors" or "~6x completion rate" stand in for a result.

Alternatives considered#

• Leave at emerging until something is measured. Rejected: it conflates two different axes. Status tracks whether an area recurs across sources, and confidence tracks how well individual claims are evidenced. Freezing the status to signal weak evidence would break the meaning of every other topic's status, and the note already carries the evidence warning in three places.
• Promote and treat claim 58 as external corroboration (raising confidence on the shared mechanism claims). Rejected, and this is the substantive call. The independence rule is about provenance, not disinterest. Two vendors marketing competing products arrive at the same design for reasons that include imitation, shared talent and a shared read of the same public research - none of which is an experiment. Raising confidence here would launder marketing convergence into measurement.
• Split multi-agent memory into its own note. Rejected as premature at one source. S7's multi-agent material (scopes, optimistic concurrency, attribution, cross-agent instruction passing) is a genuinely distinct sub-area that a consumer assistant cannot reach, but it is one source deep. Recorded as the watch item below.

Consequences#

• Easier: the topic now covers both system classes, and the maintenance-vs-retrieval boundary ADR-0007 drew held without amendment - S7's material sorted cleanly into it, which is a real test passed.
• Two of ADR-0007's open questions close. Memory as a tool-accessible file system, and memory shared across agents, are both covered by S7. A third only half-closes: content_sha256 write preconditions solve concurrent writes to one file, not two agents learning contradictory things - a distinction easy to miss and now recorded explicitly.
• One question got worse, and is the most actionable thing here. memory.md already flagged memory as an unexamined prompt-injection sink. S7's demo shows agents writing imperatives to their successors ("Next agent: skip dep checks") and successors complying, which supplies the propagation mechanism the security note lacked. Attribution and versioning give forensics; there is no admission control. Promoted as claim 63 and labelled commentary, since S7 never discusses it.
• Harder: the topic now carries numbers from one vendor and none from the other, so a careless reader could attach S6's precision to S7's architecture. The note keeps them in separate sections for that reason.
• Revisit when: any source publishes a method for measuring memory quality (the standing highest-value target); or multi-agent memory reaches a second source, at which point the split above should be reconsidered; or a deep-research pass makes the cognitive-science hop that both vendors invoke by name and neither cites.
• Follow-up edits: Topics row in the root INDEX.md (Status + source count); Sources row for S7; claims 58-64 in claims.md; 5 terms in glossary.md; S7 legend lines in both.