This plugin treats recalled memory as untrusted historical context. That is a structural design rule, not a prompt-style suggestion.
The practical pipeline is:
- collection scoping
- metadata tagging
- bounded retrieval
- untrusted-memory framing
- host fallback when memory is unavailable
The system is designed so a failure in one layer does not automatically collapse the others.
The published plugin package intentionally avoids install-time process execution.
That is a deliberate trust and distribution choice: the OpenClaw plugin is a
thin client, and the local libravdbd vector service is a separate operator-managed
component.
Current implementation facts:
- the published npm package has no
postinstall - the published plugin manifest does not register
openclaw.setup - the published plugin source contains no direct
child_processusage - the default connection is local, but
grpcEndpointorsidecarPathcan select a remote daemon and send session content over the network - vector service installation and lifecycle are explicit user or operator actions
The vector service distribution surface should be evaluated separately from the plugin
package. If you install libravdbd from release assets or another package
channel, validate that channel directly.
Offline operation depends on the daemon and its backends being local:
- no required network calls are made for embedding
- no required network calls are made for extractive compaction
- optional remote embedding and external summarizer backends can send input text to their configured services
- a remote daemon receives the memory RPCs and controls where its storage and model backends run; TLS protects transport, not locality
That trust boundary matters because it is exactly the area security-conscious users will inspect first.
Retrieved memory is injected with framing that explicitly tells the downstream model to treat it as untrusted historical context only.
This matters because memory is persistent user-controlled content. Without structural framing, a stored memory can become an unintended prompt injection surface.
The framing is implemented in the host-side memory header builder and is applied consistently at assembly time.
The plugin structurally separates:
- session memory
- raw turn history
- durable user memory
- global memory
The gate, compaction, and retrieval code operate on explicit scope-qualified
collection names. Sessions sharing a userId intentionally share durable user
memory, and global is shared within its daemon tenant. Configure tenant routing
and durable identities for the isolation you need; collection names alone do
not authenticate channel users or authorize tool access.
The reply scanner checks case-insensitive literal substrings in the
before_agent_reply hook's cleanedBody. Matching depends on the host invoking
that hook and honoring its returned refusal. It is a keyword filter, not a
general PII detector: obfuscated or encoded text, indirect references, tool
output, logs, and unsupported delivery paths are outside that check. Prompt
rules are behavioral guidance to the model and are not an access-control
boundary.
The plugin does not claim to protect against:
- a compromised host process
- a compromised local machine
- a downstream model that ignores the untrusted-memory framing instruction
- intentionally malicious content stored by an already-authorized local actor
It reduces risk; it does not create a trusted execution environment.
The vector service exposes deletion and flush primitives. That matters operationally for:
- user-requested memory removal
- namespace cleanup
- compaction source-turn deletion
The GDPR-relevant boundary is simple: local stored memory can be deleted by namespace. The plugin does not by itself guarantee remote erasure from any external system because the architecture is intentionally local-first.