Research Pulse — 2026-06-07-memclaw

# Research Pulse — 2026-06-07 (MemClaw)
_Brown Biotech tooling track · manual curation by Demis · ⚠️ marketing claim flagged_

## TL;DR
[MemClaw](https://memclaw.net) (Caura.ai) markets itself as "governed shared memory for AI agent fleets,"
framing it as Karpathy-Loop + Hyperagents infrastructure. **The pattern is real, the product is vendor-locked,
and one key marketing claim is misleading.** Brown Biotech: skip paid, watch the *primitives*, implement
locally in Hermes/PRISM where they fit.

### What MemClaw claims
_Brown Biotech relevance: multi-agent memory + contradiction detection is a real gap in our stack_

- [MemClaw website](https://memclaw.net) — Free / Pro / Enterprise tiers (pricing behind signup)
- [MemClaw blog — "The Karpathy Loop" post](https://memclaw.net) — frames the product as Hyper-Agent-generation infrastructure
- The Karpathy Loop = `karpathy/autoresearch` (one file, one metric, one time budget)
- Claims 4 tools: `memclaw_write` · `memclaw_recall` · `memclaw_insights` · `memclaw_evolve`
- Lifecycle: `active → pending → confirmed → outdated → conflicted → archived`
- MCP-server delivery (drop-in for Claude Code / Hermes / etc.)

### Verification (and a flag)
_Brown Biotech relevance: vendor marketing must be fact-checked before any paid commitment_

| Claim | Verified? | Reality |
|---|---|---|
| `karpathy/autoresearch` is real | ✅ | 85,432 stars, last push 2026-06-07 — confirmed |
| MemClaw site is live | ✅ | HTTP 200, has Free / Pro / Enterprise pages |
| **"Meta's Hyperagents research"** | ❌ **MISLEADING** | arxiv 2603.19461 = "Hyperagents" by **Jenny Zhang, Bingchen Zhao, Wannan Yang, Jakob Foerster, Jeff Clune, Minqi Jiang, Sam Devlin, Tatiana Shavrina** (Oxford + DeepMind + Microsoft + an independent). About the **Darwin Gödel Machine**, not Meta's research. |
| MemClaw implements RDF triple + vector + BM25 + permissions | Plausible (standard stack) | Marketing copy says so; no public source |
| MemClaw integrates as MCP server | Plausible | Standard pattern; verifiable on signup |
| Pricing transparency | ❌ | Free tier exists; Pro/Enterprise require signup |

**Why the Meta framing matters:** the article uses "Meta Hyperagents" + "Karpathy" as the two pillars
of inevitability — "two independent groups, same conclusion: governed memory is the missing primitive."
But the Hyperagents paper is *not* a Meta publication, and the Darwin Gödel Machine is *not* about
multi-agent shared memory (it's about self-modifying code). The actual insight — agents left to self-improve
*do* converge on persistent memory as core infra — is real and is the most interesting part. The vendor
narrative is just glued on top.

### The real primitive set (what to steal, not buy)
_Brown Biotech relevance: these 5 patterns are the actual infrastructure gap; we can build them locally_

1. **Hybrid recall (vector + BM25)** — vector finds semantic neighbors, BM25 catches exact matches (hyperparameter
   values, gene IDs, sample IDs). Composite ranking. → Hermes already has this in spirit; PRISM-RAG has
   turbovec + BM25 if wired together.
2. **Entity extraction at write-time** — every write extracts (model, hyperparam, metric, delta) and updates
   a knowledge graph. → Missing in our current Hermes; cheap to add (LLM enrichment on intake).
3. **Contradiction detection** — flag "Agent A says X" vs. "Agent B says ¬X" before one of them wastes a
   compute night. → High value, low effort: at intake, check new findings against existing findings with
   an LLM, flag conflict.
4. **Lifecycle state machine** — `active → pending → confirmed → outdated → conflicted → archived`. →
   Hermes memory is append-only; this is a real gap. Could implement as a Notion schema column.
5. **Audit trail on read + write** — every recall / write logged with agent_id, fleet_id, timestamp. →
   We have the cron-job logs but not a unified trail. Cheaper to add now than retro-fit later.

### Brown Biotech relevance
_Brown Biotech relevance: real gap, but vendor lock-in is the wrong solution; build the primitives locally_

We already have:
- **Hermes Agent memory layer** — `~/.hermes/memories/MEMORY.md` + `MEMORY_EVENTS.jsonl` (append-only)
- **PRISM RAG** — turbovec + BM25 hybrid recall (Stage 1)
- **Cron workers** — research-pulse publisher, daily signal feed, weekly review
- **Multi-OMICS agent** — separate skill pack with its own memory
- **Notion operating hub** — already a source of truth for paid briefs + funnel

The gap MemClaw names correctly: **none of these layers share findings with each other.** A contradiction
between Multi-OMICS and research-pulse would slip through today.

### Decision: **skip paid, build the 5 primitives locally**

| Track | What | Effort | Where |
|---|---|---|---|
| 1 | Add entity-extraction to research-pulse writes | 0.5 day | `brown_biotech_research_pulse_publisher.py` |
| 2 | Contradiction check at intake (LLM, ~5s, async) | 1 day | new skill: `~/.claude/skills/contradiction-check/` |
| 3 | Lifecycle column in Notion for research findings | 0.5 day | Notion schema update |
| 4 | Unified audit trail (read + write log) | 1 day | `~/.hermes/audit/` append-only JSONL |
| 5 | Multi-OMICS ↔ research-pulse share via Notion + LLM cross-ref | 1 day | cron worker |

**Total: ~4 days of work, zero vendor lock-in.** Re-evaluate MemClaw Pro in 90 days *only* if our local
implementation breaks down at scale (multi-tenant, RBAC, >10 agents).

**Skipped / deferred:**
- MemClaw Pro / Enterprise paid tier — too early, no multi-tenant need
- MemClaw Free pilot — also deferred; we can build the primitives faster than we can audit their free tier
- `memclaw_evolve` (auto-weight-adjustment) — premature; we have <2 agents that need to share findings
- Trust tiers / RBAC — overkill for solo founder + cron workers

### Verification commands
```
curl -sI https://memclaw.net → HTTP 200
arxiv 2603.19461 → Zhang/Zhao/Foerster/Clune et al. (not Meta)
karpathy/autoresearch → 85K stars, last push 2026-06-07
```

**Pattern source (read-only):** the MemClaw blog post, archived at
`~/.hermes/articles/2026-06-07-memclaw-karpathy-loop.html` (to save if not already).