GBrain sidecar experimental lane¶
Status: experimental rollout surface.
This lane is not enabled by default and does not carry stability guarantees yet. Use it as an experimental integration surface when you want to test whether GBrain can help with read-only Pack lookup or one tightly bounded helper lane.
Plain-language summary:
- Phase 1 = read-only lookup from GBrain during Pack assembly
- Phase 2 = restricted background jobs for one deterministic helper family
- Phase 3 = gated refresh apply through the existing governor path
Verdict¶
openclaw-mem stays the memory governor and ContextPack owner.
gbrain is consulted as an external retrieval/helper substrate, not promoted into a second truth owner.
Relation to gbrainMirror¶
This experimental sidecar is not the same thing as gbrainMirror.
gbrain sidecar= optional experimental lookup + bounded helper bridge during Pack/maintenance workflowsgbrainMirror= optional mem-engine write-through mirror that copies successfulmemory_storewrites into a dedicated GBrain import root
They can coexist, but they solve different problems.
If you are looking for engine-side write mirroring, read docs/mem-engine.md and extensions/openclaw-mem-engine/README.md.
Phase 1 shipped here¶
Read-only consult adapter:
openclaw-mem pack --use-gbrain onopenclaw-mem gbrain-sidecar consult --query ...
Guardrails:
- baseline
bundle_textandcontext_packstay schema-stable - gbrain output is normalized into a source-labeled additive payload
- trace receipts expose
trace.extensions.gbrain - failures are fail-open and do not break Pack
Additive payloads:
gbrain: normalized consult receiptbundle_text_with_gbrain: optional combined text for operator-controlled injection
Phase 2 shipped here¶
Bounded Minions bridge:
openclaw-mem gbrain-sidecar jobs-smokeopenclaw-mem gbrain-sidecar jobs-submitopenclaw-mem gbrain-sidecar jobs-listopenclaw-mem gbrain-sidecar jobs-retry
Current lane is intentionally narrow:
- allowed job family:
embed
Runtime note:
- durable worker execution depends on a PostgreSQL-backed GBrain engine
- pglite can still support bounded inline work, but not the persistent gbrain jobs work daemon posture
Why this narrow:
- proves the queue bridge without giving gbrain broad execution authority
- keeps phase 2 on deterministic helper work
- widens only after receipts show net value
Phase 3 shipped here¶
Governed refresh canary:
openclaw-mem gbrain-sidecar recommend-refreshopenclaw-mem optimize governor-review --approve-refreshopenclaw-mem gbrain-sidecar refresh-canary
Current posture:
- recommendation packet is additive and read-only
- governor packet is still mandatory
- canary defaults to dry-run
- actual apply is bounded to one
refresh_cardcandidate per run - receipts and rollback artifact are always written
Why this shape:
- proves the governed write bridge without opening broad mutation authority
- reuses the existing governor ladder instead of inventing a bypass lane
- keeps gbrain in the evidence role while
openclaw-memremains writer-of-record
Operator notes¶
- default binary:
gbrain - override with
--gbrain-binwhen needed - consult timeout and job timeout are explicit flags
- if
gbrainis missing or slow, Pack degrades to baseline behavior
Repo-local operator card:
Non-goals in this slice¶
- direct writes from gbrain into Store
ContextPackschema changes tied to gbrain internals- broad unrestricted job submission
- treating gbrain as truth owner