Memory and dreaming

After a year of trying special-purpose memory tools, the production recipe described at the conference is plain: markdown files in a directory, agents searching them with ordinary tools, autonomy on writes, a fast index. What made it work at fleet scale were the guardrails around it, versioning with provenance (which session motivated this change, and who made it), concurrency control on writes, and permission tiers from read-only shared knowledge down to per-agent scratch space.

Two structural limits: split focus (an agent doing a task and curating memory simultaneously does both worse), and a visibility ceiling, a session only sees itself, so a failure pattern that recurs across the fleet is invisible to every individual agent. The fix is a second-order process that runs out of band, with its own budget.

An orchestrator deploys sub-agents over a batch of session transcripts, tool calls included, not just chat, and looks for prevalent patterns: repeated failures, missing knowledge, misconfigured tools. It then proposes individual memory edits, each citing the transcripts where the pattern occurred and how often it occurred. A human accepts or rejects each proposal. Nothing lands unreviewed. One worked example: every session failing the same way traced to a tool configured in radians instead of degrees, a fix visible only in transcripts.

Capture-then-distill pipelines that hash-chain sessions and mine failures and successes into draft skills a human reads before shipping. Reviewed, confidence-scored knowledge units that graduate from personal to team to shared, with agents validating a unit before acting on it and confirming or flagging it after. Promotion-based curation: capture everything privately, promote the rare dense result outward. And attributed mistake diaries, replayed as regression tests, one team turned a failure recurring in two-thirds of runs into a consistent pass by loading the lessons.

Per-project memory lives as markdown with a one-line-per-entry index, the filesystem recipe, arrived at independently. The operating rule "refresh, don't remember" re-reads state every cycle rather than trusting recall, which is the visibility-ceiling answer applied at cycle scale.

The registers give fleet learning a durable, citable shape: per-entry record files, append-only, tamper-evident, linkable from anywhere. The conference's mistake diaries are the closest analogue, without the deletion detection.

The production implementations hand-roll hash-based compare-and-swap for concurrent memory writes. Assay's memory already lives in git, where a rejected stale write is the compare-and-swap, history is provenance for free, and a proposal arriving as a pull request inherits both the accept/reject decision and the human merge gate.

Assay's design for the consolidation pass, out-of-band transcript mining proposing evidence-cited memory updates through review, predates the conference but has not been built. What the talks changed is its status: three independent reference implementations now exist, which converts it from a research bet into a build item. Two requirements the references added to the design: keep attribution through distillation (the rendered rule must cite the incident that paid for it), and treat transcripts as untrusted input. They were mined from sessions that handled untrusted content, so proposals derived from them inherit that provenance.

8 of 41 talks shown, the ones that reach the most concepts in this theme. Every sighting, per talk, is on the concept pages above.