Show HN: Vestige – A queryless memory layer for multi agent coding fleets

I'll pull the GitHub branch and HN thread so the paragraphs stick to real product and benchmark details only.Pulling the Silent Rotation write-up and Show HN notes for exact benchmark claims.**Vestige**, a local-first memory layer for multi-agent coding fleets by Sam Vallad (samvallad33), landed on Hacker News as a Show HN tied to its **Silent Rotation** benchmark rather than a feature list. The linked artifact is a full experiment tree with harness, results, and 246 raw agent transcripts. On the board it was still early: 2 points and 3 comments. The claim under test is narrow and concrete: when three agents co-edit one failing end-to-end test in a TypeScript monorepo, and the live signing key id lives only in memory (randomized per trial from a 50-key keyring, never in a readable file), does the memory backend make the fleet converge on the true key—or on a planted decoy.

The setup runs seven arms side by side: a no-memory control, dense cosine RAG (nomic-embed-text), Vestige, mem0, supermemory, hindsight, and Zep/Graphiti. A trial passes only if the merge is green, an independent production replay passes, and the correct key is used. Across 6 models and 25 trials, no-memory fleets never converged correct (0/25) and marched to the wrong answer 21/25 times. Dense cosine RAG got 4/23 correct and 12/23 converged wrong. Vestige got 20/23 correct, 0/23 converged wrong, and 3/23 splits. On the same eight-document corpus, symptom queries from real transcripts bury the causal memory at rank 7 of 8 under both BM25 and dense cosine, while the decoy (a closed staging ticket that looks like a token-verify fix) ranks first. Vestige’s path is queryless: each agent anchors on the failure event and walks a causal backfill once, so all three agents see the same answer instead of three different retrievals from three different phrasings.

For builders running multi-agent fleets, the failure mode that matters is not “retrieval score was low.” It is asymmetric belief: two agents can hold the truth, one can hold the decoy, and the merge still fails—or worse, the fleet can agree unanimously on the wrong key, go green, and void production. Mean recall@k and nDCG hide that, because fleet safety is a conjunction over agents, not an average over queries. If each agent has a 90% chance of the right fact, a three-agent fleet is already at about 73% all-correct; at ten agents it is a coin flip. That is why a layer that removes query variance and makes supersession explicit is an engineering concern, not a demo trick.

The competitive frame is the open agent-memory stack people already wire into fleets: dense RAG, mem0, supermemory, hindsight (semantic + BM25 + cross-encoder), and Zep/Graphiti temporal graphs. Silent Rotation holds model, repo, task, tool budget, and production oracle fixed and varies only the memory backend. Vestige itself is a Rust MCP server meant to sit under Claude Code, Codex, Cursor, and other MCP clients without a cloud memory hop. The author is explicit that dense cosine tied Vestige on the Kimi K3 trials and was cheaper there, so this is not a “beats RAG on outcomes” victory lap. The separating signal in the published numbers is stability across agents and re-runs when query wording drifts.

Practical takeaway: if you ship multi-agent coding, instrument for correct convergence, retrieval consistency across agents on the same incident, and production-oracle pass rate—not only single-agent recall. Re-read the transcripts and the two-second BM25 baseline before you trust a summary table. What to watch next is the stated gap: the causal edge is hand-authored in the harness (traversal of an existing link, not automatic discovery from prose), the corpus is only eight documents, mem0/Zep cells had to be re-run after harness flush bugs, and there is still no ablation that separates event anchoring from causal walk from temporal validity. Until those land, treat Silent Rotation as a strong diagnostic of access primitives under fleet coordination, not as proof that any one product wins at production scale.