Memory as a filesystem

The arc runs from a single instruction file that was unreasonably effective and grew unboundedly, through in-band memory tools, through skills with progressive disclosure — front matter as the spine on the shelf, the body pulled only when needed — and lands on memory as a plain filesystem. The stated advice is to avoid being opinionated about special memory tools: let the store grow large, give it fast index and search, and give agents autonomy on writes. Retrieval infrastructure should be justified by measured recall failures, not by file-count anxiety. The strongest data point is someone running 1,200–1,500 large unstructured research files with no vector database and no RAG, who added a vector-plus-keyword-plus-graph layer and reported not needing it at that size.

The refinement that matters is layering. Principles and taste belong on a separate writable plane from facts and state, because the two have different lifetimes and different editors. One worked example keeps a queue for inputs, a partially discarded store of raw material, and a layered wiki separating customer insight, features, product principles and technical context — the principles layer being where taste and coherence live. The same example adds a role nobody else names: an agent whose job is refactoring the knowledge base as the organization and its constructs shift.

There is a precise technical reading of the second-brain metaphor. Weights are capacity-bounded and compress what they learn; artifacts are unbounded and queryable. Whatever cannot fit in the model lands in the record of work — experiment logs, failure reports, incident databases — which then optionally becomes the dataset that updates the weights. The corollary is organizational rather than technical: a record of every experiment, failures included, is what stops institutional knowledge walking out when people resign.

Also: Obsidian, OpenClaw, G-Brain, ZeroEntropy, the Omi wearable, Granola; the MIT-licensed, pip-installable Lab API experiment tracker.

Assay's memory plane is already per-project memory files behind an index, versioned in git, alongside a per-entry findings corpus and the stream registers. The 1,200–1,500-file result is direct evidence for the choice made here by instinct: keep memory as markdown plus an index, and spend the complexity budget on curation rather than retrieval infrastructure. The capital-markets and incident-database sources corroborate the findings design from a different direction — a failure record is organizational memory precisely because it outlives the people who produced it.

The memory index currently mixes durable rulings with ephemeral operational state: a standing decision and a tool-version workaround sit in the same list with the same lifetime. Both the layered-wiki example and the principles-versus-state split argue for separating them, and the payoff is concrete — a consolidation pass could expire stale state without touching rulings, which today it cannot do safely.

Nothing here refactors the knowledge base as Assay's own constructs shift — streams get re-homed, verbs get retired, roles get renamed. The proposed gap-closer is a scheduled knowledge-reorganization pass that files its changes as pull requests, on the same human accept-or-reject footing as any other machine-proposed edit.

• Where the no-RAG ceiling actually is. The 1,200–1,500-file figure is the only measured data point; the next scan should look for someone reporting the corpus size at which agent search stops coping.
• Whether the principles/state layering shows up as a convention anyone names, or stays an each-house-invents-it detail.
• Whether an organizational-agent-style knowledge refactor pass gets built anywhere, and what it does about churn it cannot classify.