Repository navigation
algorithm.lightweightMemory.enabled: true does not actually skip evolution pipeline (reward/L2/L3/skill still run) #2063
Description
Environment
@memtensor/memos-local-pluginv2.0.8- Adapter: opencode
- OS: macOS
Config
algorithm:
lightweightMemory:
enabled: trueExpected behavior
The schema comment says:
When enabled, the runtime skips task/reward/L2/L3/skill evolution and keeps only summarize + embedding + retrieval filter.
Expected: skill_generate / world_model_generate / policy_evolve / policy_generate stop triggering LLM calls.
Actual behavior
All evolution calls continue firing regardless of the flag. Confirmed via api_logs table:
SELECT tool_name, count(*), datetime(max(called_at)/1000,'unixepoch') as last
FROM api_logs
WHERE tool_name IN ('skill_generate','world_model_generate','policy_evolve')
GROUP BY tool_name;Results (with lightweightMemory.enabled: true active throughout):
| tool_name | count |
|---|---|
| skill_generate | 19,193 (failure rate 97.5%) |
| world_model_generate | 4,724 (failure rate 99.6%) |
| policy_evolve | 2,122 |
Root cause (identified)
buildPipelineSubscribers in dist/core/pipeline/deps.js unconditionally creates and attaches all subscribers (rewardRunner, L2, L3, skill) — the lightweightMode flag does not gate this:
export function buildPipelineSubscribers(deps, buses, algorithm, session) {
const captureRunner = createCaptureRunner({...})
const rewardRunner = createRewardRunner({...}) // always created
// L2, L3, skill runners created unconditionally as welllightweightMode only does an early return inside flush(), but the subscribers are already mounted on the event bus. The episode.finalized event still triggers the full evolution chain.
Additionally, recoverOpenEpisodesAsSessionEnd at startup re-emits episode.finalized for all previously closed episodes that lack meta.lightweightMemory === true, causing a large backlog of evolution calls after every bridge restart.
Suggested fix
In buildPipelineSubscribers, skip mounting reward/L2/L3/skill subscribers when lightweightMode is true:
export function buildPipelineSubscribers(deps, buses, algorithm, session) {
const lightweightMode = algorithm.lightweightMemory.enabled;
const captureRunner = createCaptureRunner({...})
if (!lightweightMode) {
// only create rewardRunner, L2, L3, skill when needed
const rewardRunner = createRewardRunner({...})
// ...
}Or alternatively, check lightweightMode at the top of each subscriber's event handler and return early before any LLM call.
Activity
Memtensor-AI commentedon Jul 8, 2026
🤖 AutoDev has picked up this issue and started working on it.
Task ID:
00ce0051970d8558Working branch:
feature/autodev-2063-20260708064514092Target branch:
latest dev* branchWorkflow: opsp (analysis → coding → testing → PR)
I will post the PR link here once done. If I need more information, I will ask in the comments.
fix(memos-local-plugin): actually skip evolution pipeline when lightw…
Memtensor-AI commentedon Jul 8, 2026
✅ AutoDev task
00ce0051970d8558completed.Summary: Fixed issue #2063 —
algorithm.lightweightMemory.enabled: truenow actually skips the evolution pipeline in @memtensor/memos-local-plugin. Two independent root causes were addressed. First,buildPipelineSubscribersinapps/memos-local-plugin/core/pipeline/deps.tsused to attach reward / L2 / L3 / skill / feedback subscribers unconditionally, so everyepisode.finalizedevent still cascaded through the LLM-heavy chain — the fix short-circuits the subscriber wiring and returns no-op stubs whosedrain/stop/detach/disposeare safe to call fromflush()/shutdown(), and whose runner methods throw with a "lightweight mode" message if reached (every real caller is already gated on the flag). Second,memory-core.tsstartup + periodic recovery paths (init,autoFinalizeStaleTasks,autoRescoreDirtyClosedEpisodes) used to re-emitepisode.finalizedfor legacy episodes lackingmeta.lightweightMemory === true, producing the >19kskill_generate/ >4.7kworld_model_generate/ >2.1kpolicy_evolvebacklog reported in the issue — those paths now close orphan / stale / dirty-closed episodes silently when the flag is on and tag their meta so subsequent restarts stay quiet. Regression coverage added attests/unit/pipeline/lightweight-mode.test.ts: three tests exercising normal-mode listener presence, lightweight-mode listener absence on capture / reward / L2 buses, and a syntheticepisode.finalizedrecovery emission that must produce zero reward / L2 / L3 / skill events. All 50 existing pipeline unit tests + the openclaw full-chain integration test still pass; TypeScript strict typecheck (tsc --noEmit) is clean. Three pre-existing test failures (startup-recovery,storage/migrator,v7-full-chain e2e) reproduce on the pristine base branch and are out of scope.Base branch:
mainBranch:
feature/autodev-2063-20260708064514092Commit:
c80d351PR: #2074
Assigned to: @hijzy
Reviewers: @hijzy
Fix #2063: algorithm.lightweightMemory.enabled: true does not actuall…
Fix MemTensor#2063: algorithm.lightweightMemory.enabled: true does no…