# Research sweep — papers from the fortnight to 2026-08-19

Scoped to what could change ClawMates: agent runtimes, sub-agent structure,
memory substrates, agent performance measurement, and provenance.

Each entry says what we did about it. "Nothing" is a legitimate outcome and is
recorded as such — a sweep where every paper is actionable is a sweep that
stopped judging.

---

## Skill-Use (2026-08-05) — *acted on*

Benchmarks whether an agent actually **uses** a skill under progressive
disclosure: it sees a name and description and must retrieve the procedure.
Three scores:

- **Trigger** — did it invoke the skill at all
- **Compliance** — did it follow the procedure
- **Boundary** — did it avoid what the skill forbids

That is exactly our skills shape, and reading it is what prompted asking
whether ours fire. The answer turned out to be structural rather than
behavioural: **they could not**. 55 of 85 bindings resolved to nothing, and the
catalogue had no delivery channel to a mission agent at all. Both are now fixed
(see [CAPABILITY-REVIEW.md](CAPABILITY-REVIEW.md)).

The measurement itself is now the obvious next piece of work, and it is worth
doing on the paper's three axes rather than as a pass/fail — the useful output
is *which* of our 53 skills are inert prose.

## MemoryLake on MemoryArena (2026-08-14) — *held as a finding*

A matched comparison: same framework, same model, same tasks, with the memory
backend the only changed component. Structured memory beat vector RAG and
long-context.

The numbers are the point. Success rates were **9/40, 12/20 and 4/20**, and
every system scored zero on some task category. A useful antidote to expecting
a memory swap to be transformative.

**What we did: nothing, deliberately.** It is the strongest argument against
adopting graph memory on enthusiasm, and the strongest argument for capturing a
baseline first.

## Harness the Memory (2026-08-15) — *held as a finding*

No memory substrate dominates. Broad retrieval helps factual QA, while
**excessive retrieval actively harms agent decision-making**.

Directly relevant to a decision we made this pass: mission prompts now carry
pinned skill *bodies*. This paper is the reason that is bounded
(`MAX_PINNED_SKILL_BYTES`) and restricted to pinned skills rather than "give the
agent everything it might need".

## D²ACCI (2026-08-18) — *the sharpest one for us*

> "End-to-end evaluation reveals that an error occurred, but not which stage
> caused it."

That sentence is this project's recurring defect class stated as a research
problem. The paper proposes **DCR**, a graded metric for whether failures stay
*localizable*.

This reframed the provenance assessment: the property to optimise is not
completeness of the record but **localizability of failure**. A system that
records everything and cannot tell you which stage broke has not solved the
problem. See [PROVENANCE-ASSESSMENT.md](PROVENANCE-ASSESSMENT.md).

---

## On clawhdf5 as the provenance layer

The specific question asked. `clawhdf5-agent` (~21k lines: typed entities,
relations including `RelationType::Causal`, BFS and spreading activation;
`provenance.rs` with `MemorySource` / `content_hash` / `session_id`) is
**already in the workspace manifest and used by no crate**. `cm-brain` exposes
`set_provenance` / `provenance` — a wired slot with **zero callers**.

So the adoption cost is lower than it looks. The assessment still recommends
against leading with it, for reasons that are about our system rather than its
quality: the `.brain` is reaped with the agent, which is the wrong lifetime for
an audit record, and it is file-local, so cross-agent questions need a second
mechanism anyway.

The honest sequencing is the one the memory papers argue for: make the record
complete and correctly retained on the path that matters, capture a baseline,
then measure whether the graph substrate earns its place.

## Not found

Nothing in the fortnight materially changes our **runtime** design (microVM
boot, vsock RPC, per-backend egress) or the sub-agent/topology model. Recording
that so the next sweep does not re-search the same ground.