graph-memory-zero（图形内存为零）

Graph Memory Zero

Mission

Deliver a reproducible graph-memory optimization outcome (not just a config diff):

• stable recall behavior
• explainable filtering semantics
• safe rollout + rollback
• observable runtime state

If user asks “达到你这套效果”, execute the full playbook below.

Load order (mandatory)

1. references/current-baseline.md (known-good baseline)
2. references/baseline-profiles.md (balanced/precision/recall profiles)
3. references/verification-playbook.md (acceptance checks)
4. references/troubleshooting.md (if any mismatch/failure)

When the user asks about install/download/distribution options, also load:

• references/install-channels.md

Reproducible rollout workflow

Phase 0 — Snapshot and schema guard

1. Run gateway.config.schema.lookup for:

- plugins.entries.graph-memory.config - plugins.entries.graph-memory.config.recallPolicy

1. Run gateway.config.get and store:

- current config snapshot - baseHash

1. Report: plugin enabled state, llm/embedding model, recall policy keys present.

Do not patch before confirming schema path exists.

Phase 1 — Normalize semantics (mem0-compatible)

Ensure these compatibility rules are explicitly explained in summary:

• threshold is mem0-style alias; legacy minScore may still exist.
• If both appear, effective threshold = max(threshold, minScore) (stricter wins).
• infer is deterministic inference/expansion; no extra LLM call.
• filters.memoryType supports fact|preference|task|event|all.
• preferenceLexicon (versioned) has higher priority than legacy preferenceKeywords.

If any rule is not represented in runtime config, patch minimal fields only.

Phase 2 — Apply profile patch (minimal mutation)

Default profile is balanced unless user requests otherwise.

Use gateway.config.patch with smallest scoped patch under:

• plugins.entries.graph-memory.config.recallPolicy

Balanced target (canonical):

• threshold: 0.62
• infer: true
• filters.memoryType: all
• preferenceLexicon.version: 2026-03-27.balance-v1
• preferenceLexicon.enabled: true
• preferenceLexicon.keywords: include EN+ZH preference words

If user asks for stronger precision or stronger recall, choose profile from references/baseline-profiles.md.

Phase 3 — Post-restart verification

After patch + restart, verify all below:

1. Effective config re-read matches intended patch.
2. gm_search debug details available (details.debug includes threshold/infer/filter summary).
3. No schema/key regression (memoryType not dropped, lexicon keys intact).
4. Query spot-checks pass (from verification playbook).

If any check fails, enter troubleshooting flow.

Phase 4 — Quality validation (must do before claiming success)

Run the query set in references/verification-playbook.md and compare:

• preference-sensitive queries
• task/event retrieval queries
• mixed-language (CN/EN) preference terms

Success criteria (minimum):

• relevant top hits improve or stay stable
• off-topic hits do not increase materially
• preference-related queries show better intent alignment

Do not claim “优化完成” without this phase.

Phase 5 — Rollback safety

Always keep rollback notes in output:

• previous values (before)
• target values (after)
• one-step revert patch path

If regression is observed, rollback immediately to previous stable profile.

Failure handling

A) Local test execution fails

If extension tests fail locally but config intent is clear:

1. Skip blocking local test path.
2. Use controlled gateway.config.patch rollout.
3. Run verification playbook.
4. Keep explicit rollback entry.

B) PowerShell path / command failed

If errors indicate missing path or command failure:

1. Validate path with Test-Path first.
2. Confirm script/CLI location and permissions.
3. Retry minimal command only after path is confirmed.

C) Version mismatch signals

If extension folder version and runtime installed version differ:

• treat as metadata mismatch
• continue config-level rollout, but report mismatch as release check item

Output contract (default reply structure)

Use this structure for user-facing summary:

1. 当前状态：enabled / model / embedding / recallPolicy
2. mem0 对齐语义：threshold-minScore、infer、memoryType、lexicon
3. 本次变更：before → after（只列关键键）
4. 验证结果：通过项 / 风险项 / 观测数据
5. 下一步建议：继续调优或保持当前
6. 回滚信息：可直接执行的 revert 说明

Keep answers concise-first, but never omit verification and rollback details.

Distribution guidance (when requested)

If user asks "how can others install this", provide at least 3 channels:

1. ClawHub registry install (online)
2. Offline package install (.skill as zip artifact)
3. Source-folder install (copy skill folder into workspace skills/)

Always include:

• required folder layout check (SKILL.md at skill root)
• post-install reload step (openclaw gateway restart)
• quick verification (skill appears in available skills and can be triggered)

Anti-patterns (forbid)

• Large full-config overwrite when only recallPolicy needs change.
• Declaring success without post-restart validation.
• Ignoring threshold/minScore conflict resolution.
• Omitting lexicon version in production summary.
• Hiding test/verification gaps.