Repository navigation
轻量模式下 turn_start 只取 Tier-2，导入的 skills / worldModels 永远不会被注入，而文档里没写 #2501
Description
轻量模式（algorithm.lightweightMemory.enabled）默认就是 true。而 retrieve.js 的 turnStartRetrieve() 在轻量分支里把计划写死成：
wantTier1: false,
wantTier2: true,
wantTier3: false,
Tier-1 是技能层、Tier-3 是世界模型层。也就是说默认安装下：
- skills 永远不会进 turn_start 的注入包；
- world models 也一样。
这条最容易踩在官方提供的导入路径上：importBundle() 允许导入自己写的 skills / worldModels（share: {scope:'public'}），
导完向量能查、FTS 能查、Viewer 里看得见、ranker 也会给分，但就是到不了模型那里，而且一句提示都没有。
等于默认安装下"导入自己的技能"这个功能在 turn_start 上是死的。
很难发现的原因是 memos_search 也看不出来：工具那条计划同样写死 wantTier1: false（注释说是为了避免重复注入技能）。
能看出真相的只有插件自己落盘的检索统计（stats.plannedTiers），或者注入包本身：
"stats": {"plannedTiers": {"tier1": false, "tier2": true, "tier3": false}, "raw": 20, ...}
文档上的缺口在于：schema 里写着轻量模式保留 "summarize + embedding + retrieval filter"，
那说的是演化链（reward / L2 / L3 / 技能结晶）；没有任何地方说检索的"计划"也被砍了，更没说这是默认值。
期望至少做到一条：文档里写清轻量模式会关掉 Tier-1/Tier-3（包括它是默认值），
或者让 Tier-1 在轻量模式下保留（它是纯存储读，不花模型调用）。
环境：@memtensor/memos-local-plugin 2.0.20，Windows 11；lightweightMemory 用默认值；skills/worldModels 经 importBundle() 导入。
Activity
akushonkamen commentedon Oct 16, 2026
Claiming this one.
Root cause confirmed on
main@ a7367d0 (identical to 2.0.20): the lightweight branch ofturnStartRetrievehard-codes the plan to{wantTier1: false, wantTier2: true, wantTier3: false}(apps/memos-local-plugin/core/retrieval/retrieve.ts:102-110). The intent-based scheduler plan would normally override this (applyPlanOverride), but the prompt-time recall path —recallTurninapps/memos-local-plugin/core/pipeline/orchestrator.ts:1259-1273, used by the DSH bridge turn-start (adapters/deepseek-harness/bridge.ts:228viacore.searchMemory({reason: "turn_start"})) — deliberately passesplan = undefined(no intent classification), so the hard-coded base always wins and imported skills never reach the injection packet. Tier-1 itself (runTier1) is a pure storage read (FTS + vector + pattern overrepos, no model calls), so arming it preserves the lightweight cost contract.Planned fix (minimal): arm
wantTier1in the lightweight turn-start base plan, widen the default limit bytier1TopK, keep Tier-3 off, and spell out the effective lightweight retrieval plan in thealgorithm.lightweightMemoryschema docs (core/config/schema.ts:174-180) so the remaining gap (world models not injected) is documented.Tests: a regression test asserting Tier-1 candidates surface in the lightweight turn-start packet (red before the fix), plus the existing retrieval/injection suites green.
Disclosure: analysis, code and tests are AI-assisted (Claude).
akushonkamen commentedon Oct 16, 2026
Fix is committed and locally verified.
Branch:
fix/plugin-lightweight-turnstart-tier1@63baf44(PR to follow from the contributor pipeline).Change: the lightweight base plan in
turnStartRetrieve(core/retrieval/retrieve.ts:102) now armswantTier1and widens the default limit bytier1TopK; Tier-3 stays off; the scheduler plan override still wins when present (e.g.personal_fact);algorithm.lightweightMemoryschema docs now spell out the effective lightweight retrieval plan. Also updatedtests/unit/retrieval/integration.test.ts:447-452, which had encoded the old no-skills behavior, and added a regression test reproducing the recall-path condition (no plan override) — red before the fix (plannedTiers {tier1:false}, no skill snippets), green after.Local verification:
npm run test:unit181 files — 1590 passed / 2 skipped;npm run test:integration6 passed;tests/e2e/v7-full-chain.e2e.test.ts2 passed / 1 skipped;npm run lint(tsc --noEmit) clean.Disclosure: code and tests AI-assisted (Claude).
akushonkamen commentedon Oct 16, 2026
PR with the fix: #2505 — arms Tier-1 in the lightweight turn-start base plan (pure storage read, no model calls), widens the default limit by
tier1TopK; Tier-3 stays off, scheduler plan override priority unchanged. Regression test included; docs gap filled in thealgorithm.lightweightMemoryschema.MechDesign618 commentedon Oct 16, 2026
谢谢确认，和我们这边的实测一致 —— 我们是在 DSH 端看到检索统计里 plannedTiers 恒为
{"tier1": false, "tier2": true, "tier3": false} 才定位到这条的。
Tier-1 打开这件事我们本地打补丁验证过：把 retrieve.js 轻量分支的 wantTier1 改成 true 之后，
注入包里立刻出现了 tier=1、refKind=skill 的条目（同一会话、前后相隔 77 秒，唯一变量就是那次重启），
所以你们的修法方向我们这边有实测支撑。
等含这个修复的版本发出来，我们可以在真实安装上跑一遍同样的判据（api_logs 里的 plannedTiers + 注入包本身）
并把结果贴回这里。Tier-3 保持关闭、只在 schema 文档里写明，我们也接受。