@alfe.ai/openclaw-memory-cloud

OpenClaw's private, per-agent cloud-memory plugin. It captures bounded conversation windows, recalls escaped context, and exposes tools for vector memory and the per-agent knowledge graph through the Alfe Agent API.

Safety model

Live capture state is isolated by the canonical OpenClaw sessionKey; there is no process-wide current session. Flushes for one session are serialized so the memory service's high-water mark cannot observe a later batch first. A strictly monotonic, owner-only epoch fences daemon restarts. If that state cannot be secured, auto-capture is disabled while recall and explicit tools remain available.

memory_forget is a two-step operation: the first call previews up to five current matches and the second must confirm those exact IDs in the same order. memory_learn accepts bounded inline text or a regular, non-symlink file under the configured runtime workspace. Service responses are normalized before they reach the model, and recalled context is escaped and capped at 16 KiB.

Development

pnpm --filter @alfe.ai/openclaw-memory-cloud lint
pnpm --filter @alfe.ai/openclaw-memory-cloud typecheck
pnpm --filter @alfe.ai/openclaw-memory-cloud test
pnpm --filter @alfe.ai/openclaw-memory-cloud build

The package publishes ESM and CJS entrypoints. The OpenClaw loader uses the default-only ./plugin export. Keep the tool catalog and configuration bounds aligned with openclaw.plugin.json and the memory service's agent routes.

Part of Alfe. See the documentation for platform setup.