OpenClaw Memory LanceDB Cuts Startup Memory

OpenClaw merged a focused Memory LanceDB performance improvement tonight in PR #132830, titled improve(memory-lancedb): reduce plugin startup memory. The headline result is substantial: a controlled production-loader trace showed median peak RSS dropping from 674.4 MiB to 404.3 MiB.

That is about 270 MiB less memory during plugin startup, or a 40 percent reduction in the measured path.

What Changed

The selected, enabled bundled memory-lancedb plugin was importing broad plugin SDK barrels during production loading. The PR says that, before any database, embedding request, automatic capture, recall, or dreaming work had started, the cold-process trace reached a median peak RSS of 674.4 MiB.

The repair is intentionally narrow. Memory LanceDB now imports agent-selection helpers from the focused agent-scope-runtime entrypoint, and it expresses two existing numeric TypeBox schemas directly instead of pulling them from a broader channel-actions path.

The PR states that the underlying agent-scope functions, validation constraints, tool handlers, storage behavior, and security checks are unchanged.

Why It Matters

Memory plugins are often part of long-running personal-agent setups. Startup memory overhead affects small VPS instances, local machines, home servers, and multi-agent hosts where several capabilities load together.

Reducing idle startup weight without changing the feature surface is especially useful for operators who want memory available but do not want it to dominate the process before any actual recall or storage work begins.

The change also makes plugin startup more honest. A memory backend should spend memory on database and embedding work when those features are used, not on importing broad SDK surfaces just to register its entrypoint.

Evidence From The PR

The PR includes a reproducible production-loader benchmark using Node v24.19.0 and the built OpenClaw runtime graph. The before revision measured three runs with a median maximum RSS of 690,600 kB. The after revision measured a primary three-run median of 414,036 kB, with two additional confirmation runs at 413,988 kB.

The same trace confirmed that the plugin still loaded from its built runtime artifact and registered all three memory tools: memory_recall, memory_store, and memory_forget.

Validation covered both behavior and guardrails:

• extensions/memory-lancedb/startup-imports.test.ts and extensions/memory-lancedb/index.test.ts passed 147 tests.
• pnpm build passed.
• pnpm check:changed passed, including extension typechecks, lint, plugin contracts, boundary checks, storage guards, and import-cycle checks.
• A security diff review found no issues in the focused entrypoint change.

The Takeaway

This is a good example of the kind of performance work that matters in agent runtimes. It does not add a new button or command, but it lowers the cost of keeping a memory backend ready.

For Memory LanceDB users, the important promise is that the plugin should start lighter while keeping the same tools and constraints. For OpenClaw as a platform, it is another step toward making always-on agent stacks more practical on constrained hosts.