Imported from Sundried-calomel/memory-wuxian (SKILL.md). Install upstream with npx skills add Sundried-calomel/memory-wuxian. Copyright stays with the author.

Memory無限

Build effectively unbounded, retrievable conversation memory from immutable source records and persistent indexes.

Core invariants

1. Persist the exact source message before allowing runtime compression.
2. Treat raw records as append-only historical authority.
3. Use summaries only as routing indexes; verify historical claims against raw text.
4. Build deterministic Level-1 indexes after the configured completed-round or visible-character threshold, whichever occurs first.
5. Generate each higher level after a configured number of ungrouped child summaries.
6. Persist every summary level and every parent-child relationship.
7. Record only explicit source information. Do not infer hidden preferences, motives, or importance.
8. Keep runtime compression separate from permanent memory.
9. Verify SHA-256 source integrity before summary ingestion.
10. Rebuild only derived state and indexes; never repair integrity failures by rewriting history.
11. When Codex integration is enabled, import user-visible dialogue, lightweight tool activity visible in the task timeline, and successful structured file-change events. Preserve file paths, change types, line ranges, addition/deletion counts, and exact unified diffs. Exclude general tool output and hidden reasoning.
12. Complete the primary archive write before creating its external backup snapshot.
13. Maintain one complete transcript per conversation; never place records from different conversation IDs in the same transcript.
14. Use the native event-driven collector for high-frequency Codex capture on macOS and Windows; keep Python outside the continuous capture loop.
15. Preserve transaction consistency by holding memory/.locks/archive.lock for each native event batch and Python maintenance command.
16. Keep summary source ranges, parent-child groups, and derived indexes scoped to one conversation ID.
17. Exclude native Codex subagent sessions; archive only top-level user-visible conversations.
18. Keep only the configured number of newest complete external snapshots; the default is one.
19. Keep only the configured number of newest workspace recovery backups under memory/archive/; the default is one.
20. Do not keep an AI conversation active. Let scripts detect completed-round or character thresholds, then run one ephemeral AI process only to generate the due semantic summary.
21. Treat dashboard snapshots as disposable derived caches. Render the last persisted snapshot immediately, validate and rebuild it in a background thread for automatic refresh, and reserve synchronous validation for explicit manual refresh.
22. Keep each node's local archive exclusively writable by that node. Store imported peer history only in read-only replicas under the federation cache.
23. Qualify federated message, conversation, and summary identities by origin node. Never merge remote records into local counters or authoritative raw files.
24. Export only locally originated artifacts. Verify artifact SHA-256, event-sequence continuity, and predecessor bundle SHA-256 before committing an import.
25. Treat .mwxb as a compressed integrity-checked container, not as an encrypted or signed message. Transfer it only through SSH or another trusted channel.
26. Keep federation identity separate from OpenAI sessions and exclude reconstructible peer replicas from primary-archive desktop backups.
27. Keep SSH and encrypted cloud-folder exchange as parallel transports over the same .mwxb import contract.
28. Sign every cloud-bound delta with the origin device identity and encrypt it to the target device before it enters a synchronized folder.
29. Keep cloud private keys on their owning devices and never store cloud-account credentials in Memory無限.
30. Run cloud synchronization as a low-frequency, short-lived, model-free task. Do not place cloud polling in the native collector.
31. Treat ChatGPT export import as an explicit local operation. Never upload the selected export, and label the adapter experimental until a real official user export has been validated.
32. Keep README.md, README.zh-CN.md, and README.ja.md semantically synchronized whenever documented features, installation, commands, limitations, privacy boundaries, or release behavior change.
33. Record explicit operational-rule changes as append-only Level-1 policy events. Require exact prior-statement linkage before a revision, withdrawal, or reaffirmation changes current validity.
34. Use current-policy retrieval for rules or strategies that may have changed. Recency alone must never supersede an earlier policy.
35. Treat the desktop dashboard as a required release artifact. Every installer or update that changes dashboard code, runtime paths, or installation behavior must replace the platform launcher, preserve the active archive, bind the launcher version to the package version, and pass a live post-install self-check before the release is considered complete.
36. Keep local-hash-v1 as the no-download semantic default. The optional multilingual-e5-small provider must use the pinned installer, immutable model revision, exact artifact SHA-256 verification, an isolated runtime, offline inference, and disabled remote model code. Semantic vectors remain disposable derived data and every returned hit must be verified against raw history.
37. Rebuild the Windows native-dashboard shortcut after every installer upgrade, using the preserved active archive root and the Python runtime validated by the current bootstrap.
38. Persist top-level Codex token_count telemetry in a separate derived per-conversation ledger. Call it Codex-reported model usage, not billing usage. Treat cached input and reasoning output as included subfields, detect cumulative-counter resets, exclude subagents, and never place telemetry in raw dialogue or semantic summaries.
39. Keep Environment Registry authority, locks, cursors, staging, and receipts independent from the 1.x conversation archive. Environment operations must never rewrite raw dialogue, summaries, or local archive ownership.
40. Represent global Rules, project Rules, global Skills, and project Skills as immutable, content-addressed revisions with explicit node-local bindings. Synchronize managed rule blocks only; preserve all bytes outside those blocks.
41. Exchange environment revisions through a separate signed and target-encrypted environment-v1 stream. Each device keeps its own writable local state and receives remote conversation history only as read-only replicas.
42. Let the five-minute cloud task validate and stage incoming Environment updates without AI. No-change must create no decision, receipt, backup, or model call.
43. Never auto-install a Skill or project-scoped artifact. Divergence, identity changes, permission expansion, persistent-component expansion, and incompatible runtimes require explicit review and fail closed.
44. Install only a registered immutable revision through a verified binding.
45. Keep mechanical maintenance model-free and persistent. Use stable idempotency keys, leases, bounded retries, and quarantine; promote semantic work only after a complete dialogue boundary. Each job receives one explicitly leased one-shot AI worker attempt. One five-minute maintenance batch may overlap at most three model calls, while source verification, archive ingestion, parent-job creation, and state/index writes remain serialized through their existing locks.
46. Keep exact-byte content storage in the removable shadow path. Require ordered closed manifests, per-file length and SHA-256, contiguous per-stream checkpoints, exact reconstruction verification, and preview-first writes. Never treat shadow data as replacement authority. Validate package contents, platform and runtime contracts, preserve a rollback object before mutation, atomically switch, self-check, and append a receipt.
47. Promote reusable project capability to global scope only through a separate evidence-bearing proposal, complete platform matrix, and explicit approval.
48. Transport governance-insight proposals only as immutable, source-bound evidence. Keep imported proposals in read-only peer replicas; transfer, repetition, or arrival from another device never constitutes semantic review, acceptance, Rule registration, or Skill installation.
49. Keep governance-AI orchestration disabled until explicitly enabled. Five-minute checks are model-free; invoke at most one ephemeral Codex worker only when a bounded compatible batch is due.
50. Keep product work on its source device and run cross-device governance classification only on the explicitly configured coordinator.
51. Treat every AI result as a schema-validated, reviewable draft. It must never accept a Rule, install a Skill, remediate a product, rewrite history, or mark itself human-reviewed.
52. Validate evidence hashes before invocation. Retry a failed item at most twice, then isolate it without blocking unrelated queue work.
53. Assign every exported Environment Registry item a stable identity derived from its transaction, operation, and immutable object identity. Batch registration must export every artifact and project exactly once.
54. Parse Skill metadata with a safe full YAML loader. Permit legal nested mappings, lists, and block scalars; reject unsafe tags and duplicate keys. Project registrations received from peers remain read-only replicas and never create or activate a local project automatically.
55. The native dashboard must bind an operating-system-assigned loopback port and open the server's actual port. Never share or assume another local application's fixed port.
56. Treat PRODUCT_ARCHITECTURE.md as the canonical module-boundary owner and docs/module-architecture.json as its machine-readable source-ownership registry. Register each new production file under exactly one owner and pass scripts/check_architecture_contract.py for every product change. Unowned files, overlapping owners, and prohibited dependencies fail closed.
57. Treat a visible cloud-provider placeholder that is not locally readable as transient. Trigger bounded hydration before decrypting it and never quarantine a Files On-Demand availability error as cryptographic damage.
58. Recover an overlapping environment-v1 range only after every persisted prefix event matches exactly. Prefer the widest newest valid candidate for one expected start, preserve exact replay for lost acknowledgements, and keep archive and Environment status histories independent.
59. Preserve stable executable entry paths in macOS background definitions. Never resolve a Homebrew or managed-runtime symlink into a version-specific installation directory, because a runtime upgrade would create a new macOS privacy identity and can trigger repeated Desktop or Documents prompts.
60. Apply an existing macOS installation only through the user-space candidate transaction. Probe exact-message capture in an isolated archive before cutover; acquire the shared archive lock, require zero native recovery debt, and stop the previous collector at that idle boundary before replacing files. After cutover verify a new live collector and the current dashboard; on any handoff or post-switch failure restore the previous Skill, plist, and collector.
61. Publish collector telemetry on every monitoring interval, including idle intervals. Keep source and archive watermarks separate, expose stale telemetry and archive lag, and never infer archive freshness only from a running PID.
62. Before generating a report for a historical cutoff, run the deterministic archive-waterline preflight. If retained source bytes through the cutoff are not covered by persisted cursors, stop or perform the explicit bounded backfill and verify again before using Memory無限 evidence.
63. Synchronize semantic capability through an immutable global-runtime-contract, not by copying a platform virtual environment. Pin the model revision, artifact hashes, runtime packages, embedding interface, and platform-neutral installer entry. Receiving or accepting a contract never installs or downloads it; local realization requires an explicit reviewed --apply, and semantic indexes remain device-local.
64. On an existing macOS installation, routine stable-release updates must verify the PKG checksum, extract only its Skill payload in a temporary directory, and invoke install_macos_transaction.py. Do not open the platform installer or request administrator credentials unless this is a first install, recovery, or declared privileged-component migration.
65. Before assigning or implementing v2.6 or later, read references/version-roadmap-v2.5-to-v3.0.md. Preserve its version order, predecessor gates, non-goals, evidence matrix, and rollback contract. Personal Environment convergence is v2.10 after v2.9; continuous catch-up and bounded debt convergence is compatible v2.11 work. v3.0 is conditional, and an untagged future-version branch is not a release.
66. Use one bounded provenance-aware read service for CLI, loopback HTTP, and MCP. Keep adapter validation and confidence behavior equivalent; expose no write, deletion, pairing, installation, arbitrary path, command, or remote control operation. Verify update metadata against the pinned signer, stage verified bytes without execution until a separate version-and-hash-bound explicit approval, and keep summary-budget eligibility model-free.
67. Treat a personal Environment Profile as immutable evidence, never as installation authority. Capture only explicitly supplied global Rule files and Skill roots; omit paths, credentials, device identity, archives, conversations, indexes, models, and caches. Capture is preview-first and --apply may only create a predecessor-linked generation and current pointer. Imported generations remain read-only peer replicas. Comparison and convergence planning never invoke a Rule or Skill installer.
68. Preserve the earliest collector activation boundary across installation and upgrade. A retained top-level rollout without a completed cursor is coverage debt even when it predates the latest installer run. Recover it through bounded native streaming batches, advance the cursor only after durable append, and never rewrite partial historical records. Reconcile mechanical, semantic, and backup debt with a hidden bounded supervisor; model unavailability must defer semantic work without consuming retries. Every actual semantic prompt must satisfy both character and UTF-8 byte budgets.
69. Daily archive-volume reporting is a derived, read-only projection. Define all devices as the local node plus each currently trusted synchronized peer exactly once, use the Asia/Tokyo day boundary, and label Codex token values only from token_count telemetry. Exchange only immutable path-sanitized token-ledger revisions through federation protocol v2, retain v1 bundle read compatibility, and expose missing or stale peer telemetry instead of converting partial coverage into a zero or account-global claim.

Operating workflow

1. On Windows, run powershell -ExecutionPolicy Bypass -File scripts/bootstrap_windows.ps1 before the first archive operation. Pass -AgentsPath <workspace AGENTS.md> to install or deterministically update the canonical workspace rules. If it reports missing-runtime, rerun with -InstallMissing after user approval. Reuse Codex-bundled Python and CLI when available; do not install Rust or MSVC unless rebuilding the collector.
2. Run python3 scripts/memory_cli.py init for a new memory root. The Windows collector installer records its --archive-root as the active archive, so later CLI calls can omit --root. An explicit --root or MEMORY_WUXIAN_ROOT still overrides that pointer.
3. Append each user and assistant message with append; one user message plus its assistant response forms a completed round.
4. Let the native collector mark a summary due after 5 completed rounds or 20,000 visible characters. A character threshold reached during an answer is acted on only after that answer's final_answer closes the round.
5. Let the one-shot semantic worker generate and ingest the AI summary, then exit. The hidden maintenance batch may run at most three independent model calls concurrently, but each verified result is ingested under the existing archive and summary locks. Use make-summary-job and summary prompt for manual recovery.
6. Use retrieve for earlier topics. Let it search indexes first and raw records second. Retrieval is read-only and does not require the archive write lock; query logging is skipped automatically when the caller lacks write permission.
7. Base answers on the recovered raw segment and report the returned verification level.
8. Run heartbeat for validation and recovery. Keep count-based events as primary triggers.
9. Preview rebuild-state, rebuild-conversations, or rebuild-indexes before applying a recovery operation.
10. Use the native collector for automatic Codex import. Use sync-codex only as a manual compatibility and recovery adapter. Both paths must remain idempotent and storage-compatible. Let maintenance_supervisor.py run as the installed hidden five-minute one-shot scheduler. It reconciles all historical debt, skips quarantined items without blocking later work, and resumes deferred semantic jobs when Codex becomes available. Do not replace it with a foreground loop.
11. Use import-chatgpt for an official ChatGPT data-export ZIP, extracted directory, or conversations.json. It is incremental and idempotent, but it is not a real-time ChatGPT listener. The same experimental adapter is available under Dashboard > Settings; current automated coverage uses synthetic exports and does not constitute validation against a real user export.
12. When desktop backup is configured, confirm the returned snapshot path after each successful mutation.
13. Use backup to create a verified recovery snapshot on demand and prune snapshots beyond configured retention.
14. Before editing this Skill, refresh one replaceable workspace code backup instead of adding timestamped copies. Never place a full live archive in development outputs.
15. At the start of each user turn, run context-refresh-status. When due, load context-capsule only if its refresh_id is not already present in the active reasoning context. Capsule reads are fully read-only and require no acknowledgement. Never run ack-context-refresh as part of normal operation, never stop or request permission for an acknowledgement, and never archive a capsule as a source message. The legacy ACK command is a compatibility no-op.
16. When the user names another or historical Codex conversation and asks to continue it or restore its latest messages, run conversation-tail --title "..." --exclude-conversation-id "codex:<active-task-id>" --messages N. Resolve the title after excluding the active task and before selecting messages. Never substitute the latest conversation when the title is missing or ambiguous. When the user confirms a title-to-task relationship, persist it with register-title so later retrieval does not depend on mutable client title metadata.
17. Let the dashboard render its last successful browser-local response immediately. Serve memory/dashboard/status-snapshot.json without blocking the first paint, rebuild it from authoritative records in the background, and animate changed values when the refreshed snapshot arrives. On Windows, dashboard status reads must not create visible console subprocesses.
18. For federation, run init-node once, register only explicitly trusted peers, and use export-delta, inspect-bundle, and import-delta for offline exchange.
19. Use sync-peer only after SSH host identity is present in the local known-hosts trust store. Select posix or powershell to match the remote shell.
20. Use retrieve-global for cross-device history. Treat a peer result as verified only after its imported artifact hash has been checked.
21. Use revoke-peer to reject future imports and SSH pulls from a device. Revocation does not silently delete previously imported history.
22. Use cloud-configure, cloud-pair-export, and cloud-pair-import to prepare an explicitly selected iCloud Drive, OneDrive, or compatible synchronized directory.
23. Let users manage routine cloud synchronization from Dashboard > Settings. The cloud switch must enable or disable both transport configuration and its background scheduler, the status view must expose the configured directory and scheduler state, and the manual sync command must run one encrypted exchange pass without requiring an AI conversation.
24. Keep cloud-enable, cloud-disable, and cloud-sync as equivalent CLI and recovery controls. The scheduled task wakes every five minutes, while ordinary exports are coalesced and empty checks create no files.
25. Treat all three localized README files as one documentation contract. Update and verify English, Simplified Chinese, and Japanese in the same change.
26. On macOS, let the package installer rebuild ~/Desktop/Memory無限操作台.app from the packaged application, refresh its configuration from the current Python, Skill, and active archive paths, then verify its version, code signature, executable hash, and self-check.
27. On Windows installation or upgrade, preserve the active archive pointer and recreate Memory无限状态台.lnk by default; never retain an absolute Python path from an older Codex runtime. A bare Skill copy has no traditional installer UI, so first activation must run the supplied bootstrap and shortcut installer.
28. Let the native collector update each conversation's Token ledger incrementally. Use token-usage-backfill in preview mode before --apply when retained historical Codex rollout files need to be measured. Do not infer missing usage from archived text or claim that ChatGPT exports expose model-consumption telemetry.
29. Use environment-init, explicit root and project bindings, and environment-register to establish Environment Registry state. Preview every registration or install before --apply.
30. Let cloud-sync process archive, Environment, and project-evidence streams. Incoming Environment material is validated into staging first; use environment-incoming-status and the dashboard Environment view before an explicit acceptance or installation.
31. Use environment-profile-capture --specification <json> to preview one path-free personal Environment generation and add --apply only to persist it. Use environment-profile-compare and environment-convergence-plan for trusted peer evidence. These commands never install, bind, enable, or overwrite a Rule or Skill.
32. Build Project Evidence Packages only from an explicit bounded manifest. Preserve exact bytes and hashes, keep imported copies read-only, never activate them, and use the independent project-evidence-v1 stream so old clients can ignore it safely. Follow references/project-evidence.md. Keep Project Evidence Owners device-local and explicit. Each owner binds a closed file selection to one source root, never discovers a workspace, and never exports its source path. A five-minute model-free pass may refresh at most 20 owners; unchanged content creates no record, changed stable content creates one predecessor-linked generation, and failures remain isolated. Put explicitly selected large final deliverables in the independent project-attachment-v1 stream. Use 4 MiB exact chunks, keep each logical file at or below 256 MiB and each generation at or below 1 GiB, and never rewrite or replace the ordinary source file. Use project-attachment-sync when only attachment transfer is authorized. Treat upload, acknowledgement, and verified reconstruction as separate states; only a complete SHA-256 reconstruction may write a receipt. Follow references/project-attachments.md.
33. Use environment-conflicts and environment-promotions for current governance state. Resolve a conflict or advance a promotion only with explicit reviewer and evidence fields; never infer approval from recency or successful transfer.
34. Validate a governance insight with work-system-governor, then use environment-governance-propose to preview and persist the immutable local envelope. Use environment-governance-proposals to inspect local and peer proposals; acceptance remains a separate global Owner workflow.
35. Validate product evolution reports with work-system-governor, then use environment-product-evolution-record to preserve and exchange the immutable evidence record. Peer records stay read-only and never trigger product remediation or governance acceptance.
36. Use environment-governance-ai-discover for model-free discovery, environment-governance-ai-status for inspection, and environment-governance-ai-tick for a bounded one-shot review. Configure or enqueue only through preview-first CLI commands. Human review remains mandatory before any downstream acceptance or installation.
37. Before implementing a product change, identify its canonical owner in PRODUCT_ARCHITECTURE.md. If it adds or relocates production code, update docs/module-architecture.json first, then run the architecture contract before focused behavioral tests.
38. On macOS, run scripts/install_macos_transaction.py for an existing installation. Treat its isolated candidate probe, live PID replacement, telemetry freshness, dashboard self-check, and rollback proof as one indivisible update contract.
39. Before a time-bounded report reads Memory無限, run scripts/archive_waterline.py --cutoff <ISO-8601>. Use --backfill only for the exact lagging retained rollout files and require a final covered result.
40. During collector startup, persist any due semantic job but defer AI execution to the independent semantic-backfill scheduler. Do not hold collector readiness or update cutover open while a Codex CLI summary runs.
41. Let native capture atomically record coalescing backup debt instead of copying the complete archive inline. Let the maintenance worker create one complete snapshot for all pending mutations, and clear the debt only after that snapshot succeeds.
42. Register the bundled E5 interface with environment-register-semantic-runtime --origin-node-id <node>, let the existing environment-v1 stream transport it, and use environment-realize-semantic-runtime --apply only after explicit review on the receiving device. Use semantic-runtime-status to verify the bundled contract, registered revision, model artifacts, and local runtime.
43. Treat configuration-compile, configuration-explain, and environment-capability-status as stateless read-only diagnostics. They must not construct MemoryStore, initialize an archive, take an archive lock, alter trust, grant permissions, install capabilities, or start synchronization.
44. Keep the v2.5 effective-configuration contract closed and deterministic. Unknown or duplicate keys and invalid values fail closed. Preserve --root, MEMORY_WUXIAN_ROOT, active-root pointer, then configured-root precedence.
45. Memory-sharing scopes are design-only until a separately approved multi-user, third-party-write, partial-sharing, hosted-service, non-shareable-data, or cross-identity requirement activates the review. Do not add runtime scope fields or controls before that decision.
46. For v2.6-or-later work, report the target version, predecessor evidence, canonical owner, changed contracts, preserved invariants, test gates, and rollback path from references/version-roadmap-v2.5-to-v3.0.md before implementation or handoff.
47. Build v2.6 shadow indexes with index-generation-build, verify them with index-generation-status, and keep activation and rollback preview-first. Never use an index generation to rewrite raw history, and never activate a received generation without an explicit --apply operation.
48. Run scripts/runtime_effect_gate.py for release and post-install effect verification. Do not equate a running process, registered scheduler, created file, mocked result, or zero exit code with a working background feature. Stale waterlines, hidden fallbacks, permanent debt, incomplete backups, missing parent-summary work, and stale supervisor state fail the gate.
49. Keep native capture independent from AI execution. The collector may persist a semantic job but only the independent leased dispatcher may run it. After a Level-1 summary is ingested, deterministically enqueue any due parent summary. Bind semantic indexes to the exact raw-source snapshot and expose keyword fallback when stale. During upgrades, add only missing configuration defaults and preserve the prior bytes as rollback evidence.
50. On Windows, prefer a validated explicit installed Skill root over the process SID profile. After every install or upgrade, resolve the final desktop .lnk through Windows Shell and verify its exact launcher target, working directory, icon, empty arguments, launcher configuration, and live target. A shortcut that only exists does not prove activation.

Commands

python3 scripts/memory_cli.py init
python3 scripts/memory_cli.py append --speaker user --text "..."
python3 scripts/memory_cli.py append --speaker assistant --text "..."
python3 scripts/memory_cli.py sync-codex --session-file ~/.codex/sessions/YYYY/MM/DD/rollout-....jsonl
python3 scripts/memory_cli.py token-usage-backfill
python3 scripts/memory_cli.py token-usage-backfill --apply
python3 scripts/memory_cli.py import-chatgpt --export /path/to/chatgpt-export.zip
python3 scripts/memory_cli.py status
python3 scripts/memory_cli.py context-refresh-status
python3 scripts/memory_cli.py context-capsule
python3 scripts/memory_dashboard.py --root /path/to/archive --config /path/to/config.yaml --window
python3 scripts/memory_cli.py backup
python3 scripts/memory_cli.py make-summary-job
python3 scripts/semantic_worker.py --root memory --config config.yaml --job memory/pending/<job>.json
python3 scripts/semantic_backfill.py --root memory --config config.yaml --max-jobs 20
python3 scripts/memory_cli.py maintenance-enqueue --kind archive-health --idempotency-key health:manual
python3 scripts/memory_cli.py maintenance-status
python3 scripts/memory_cli.py maintenance-tick --maximum-jobs 20
python3 scripts/memory_cli.py maintenance-diagnostics
python3 scripts/memory_cli.py content-shadow-build --source-root /snapshot --source-id node:snapshot --file raw/a.md
python3 scripts/memory_cli.py content-shadow-build --source-root /snapshot --source-id node:snapshot --file raw/a.md --apply
python3 scripts/memory_cli.py content-shadow-status
python3 scripts/memory_cli.py content-shadow-verify --manifest-id <manifest-id> --source-root /snapshot
python3 scripts/memory_cli.py content-shadow-reconstruct --manifest-id <manifest-id> --destination /restore
python3 scripts/memory_cli.py content-shadow-reconstruct --manifest-id <manifest-id> --destination /restore --apply
python3 scripts/memory_cli.py content-transfer --manifest-id <manifest-id> --target-archive-root /target --domain archive --target-id <node> --start 0 --count 100
python3 scripts/memory_cli.py content-shadow-disable
python3 scripts/memory_cli.py content-shadow-disable --apply
python3 scripts/memory_cli.py ingest-summary --job memory/pending/<job>.json --summary-json <summary>.json
python3 scripts/memory_cli.py retrieve --query "..."
python3 scripts/memory_cli.py retrieve --query "..." --mode current-policy
python3 scripts/memory_cli.py semantic-runtime-status
python3 scripts/memory_cli.py environment-register-semantic-runtime --origin-node-id <node>
python3 scripts/memory_cli.py environment-register-semantic-runtime --origin-node-id <node> --apply
python3 scripts/memory_cli.py environment-realize-semantic-runtime
python3 scripts/memory_cli.py environment-realize-semantic-runtime --apply
python3 scripts/memory_cli.py conversation-tail --title "Codex conversation title" --exclude-conversation-id "codex:<active-task-id>" --messages 20
python3 scripts/memory_cli.py register-title --conversation-id "codex:<task-id>" --title "Confirmed title"
python3 scripts/memory_cli.py rebuild-state
python3 scripts/memory_cli.py rebuild-state --apply
python3 scripts/memory_cli.py rebuild-conversations
python3 scripts/memory_cli.py rebuild-conversations --apply
python3 scripts/memory_cli.py rebuild-indexes
python3 scripts/memory_cli.py rebuild-indexes --apply
python3 scripts/memory_cli.py index-generation-build
python3 scripts/memory_cli.py index-generation-status --generation-id <id>
python3 scripts/memory_cli.py index-generation-activate --generation-id <id>
python3 scripts/memory_cli.py index-generation-activate --generation-id <id> --apply
python3 scripts/memory_cli.py index-generation-rollback
python3 scripts/memory_cli.py index-generation-rollback --apply
python3 scripts/memory_cli.py rebuild-deterministic-indexes
python3 scripts/memory_cli.py heartbeat --check-only
python3 scripts/memory_cli.py heartbeat
python3 scripts/memory_cli.py heartbeat --repair
python3 scripts/memory_cli.py init-node --display-name "This computer"
python3 scripts/memory_cli.py add-peer --node-id <peer-node-id>
python3 scripts/memory_cli.py export-delta --output /trusted/path/update.mwxb --target-node-id <peer-node-id>
python3 scripts/memory_cli.py inspect-bundle --bundle /trusted/path/update.mwxb
python3 scripts/memory_cli.py import-delta --bundle /trusted/path/update.mwxb --expected-node-id <peer-node-id>
python3 scripts/memory_cli.py rebuild-global-index
python3 scripts/memory_cli.py retrieve-global --query "..."
python3 scripts/memory_cli.py federation-status
python3 scripts/memory_cli.py sync-peer --node-id <peer-node-id>
python3 scripts/memory_cli.py revoke-peer --node-id <peer-node-id>
python3 scripts/memory_cli.py cloud-configure --directory /path/to/synchronized/MemoryWuxianExchange
python3 scripts/memory_cli.py cloud-pair-export
python3 scripts/memory_cli.py cloud-pair-import --pairing-file /trusted/path/peer.json
python3 scripts/memory_cli.py cloud-enable
python3 scripts/memory_cli.py cloud-disable
python3 scripts/memory_cli.py cloud-sync
python3 scripts/memory_cli.py cloud-sync --force
python3 scripts/memory_cli.py cloud-status
python3 scripts/memory_cli.py environment-init
python3 scripts/memory_cli.py environment-status
python3 scripts/memory_cli.py environment-incoming-status
python3 scripts/memory_cli.py environment-process-incoming
python3 scripts/memory_cli.py environment-conflicts
python3 scripts/memory_cli.py environment-promotions
python3 scripts/memory_cli.py environment-governance-propose --proposal-json /path/to/proposal.json
python3 scripts/memory_cli.py environment-governance-proposals
python3 scripts/memory_cli.py environment-product-evolution-record --record-json /path/to/product-evolution.json
python3 scripts/memory_cli.py environment-product-evolution-records
python3 scripts/memory_cli.py environment-governance-ai-discover
python3 scripts/memory_cli.py environment-governance-ai-status
python3 scripts/memory_cli.py environment-governance-ai-enqueue --item-json /path/to/item.json
python3 scripts/memory_cli.py environment-governance-ai-configure --policy-json /path/to/policy.json
python3 scripts/memory_cli.py environment-governance-ai-tick --run-ai --maximum-batches 1
python3 scripts/memory_cli.py configuration-compile
python3 scripts/memory_cli.py configuration-explain
python3 scripts/memory_cli.py environment-capability-status
python3 scripts/memory_cli.py environment-capability-status --peer-offer /path/to/peer-offer.json
python3 scripts/install_governance_ai.py --archive-root /path/to/memory --skill-root /path/to/memory-wuxian --python-executable /path/to/python --load
scripts/build_native_collector.sh
python3 scripts/install_codex_autosync.py --archive-root /path/to/memory --load
powershell -ExecutionPolicy Bypass -File scripts/build_native_collector.ps1
powershell -ExecutionPolicy Bypass -File scripts/bootstrap_windows.ps1
python scripts/install_agent_rules.py --agents-file /path/to/workspace/AGENTS.md
python scripts/install_codex_autosync_windows.py --archive-root C:\path\to\memory --load
python scripts/install_auto_update.py --skill-root /path/to/memory-wuxian
python scripts/auto_update.py --check-only --force
python scripts/auto_update.py --approve-install --expected-version <version> --expected-sha256 <sha256>
python scripts/memory_cli.py readonly-query --query "prior decision" --mode hybrid --limit 20
python scripts/memory_cli.py readonly-http --host 127.0.0.1 --port 8766
python scripts/memory_cli.py readonly-mcp
python scripts/memory_cli.py summary-budget-status --metrics-json /path/to/metrics.json --policy-json /path/to/policy.json
powershell -ExecutionPolicy Bypass -File scripts/install_dashboard_shortcut_windows.ps1 -SkillRoot C:\path\to\memory-wuxian -ArchiveRoot C:\path\to\memory -PythonExecutable C:\path\to\python.exe

Use semantic_backfill.py for historical summary debt. It processes higher-level parent jobs before Level-1 jobs, is safe to rerun, and creates one recovery snapshot after the batch instead of copying the complete archive after every summary. Keep --max-jobs bounded for routine maintenance; --max-jobs 0 drains all due work. The configured model-call concurrency is hard-bounded to three within one batch; result ingestion remains serialized and a failed sibling does not cancel successful jobs. A clean full-recovery audit may be reused for 24 hours unless explicit recovery debt exists.

Pass --root <memory-directory> before the subcommand to use a memory archive outside this skill folder.

Load supporting material selectively

• Read implementation.md before changing storage formats, counters, summary hierarchy, retrieval behavior, state recovery, locking, privacy behavior, or client integration.
• Read schemas.md when constructing or validating raw records, summary JSON, indexes, state, or retrieval output.
• Read decisions.md before changing architectural behavior.
• Read deferred-memory-scope-design.md before proposing memory-sharing boundaries. It is a design trigger, not an implemented privacy feature.
• Read release-rehearsal.md before release claims. Run scripts/run_release_rehearsal.py; never describe unrun or evidence-free scenarios as passed. During candidate CI, run the full unittest suite once and pass its retained log with --reuse-unittest-evidence; the rehearsal then emits one hashed reference log per covered contract without executing those modules again. Feature branches use PR CI only, main uses one complete same-SHA candidate gate, and the release workflow builds installers from that proof.
• For a bounded patch, declare validation_profile: targeted-patch and an explicit affected scenario list in its version work contract, then run run_release_rehearsal.py --contract-profile. Do not restart unrelated historical suites for a small correction. Missing or expanded scope defaults to the complete candidate gate.
• Read AGENTS.md when integrating this skill into an Agent's persistent operating rules.
• Use files in templates/ as output contracts and files in prompts/ as Agent prompts.

Client integration boundary

Installing the Skill alone does not intercept Codex events. Automatic capture requires the supplied macOS LaunchAgent or Windows scheduled task. Both keep only the Rust collector alive, use immediate native filesystem events plus an adaptive 5-second, 30-second, and 5-minute metadata fallback, and share the same archive contract. They import user messages, visible assistant commentary/final answers, lightweight task-timeline tool activity, and successful structured file-change diffs from top-level sessions; they exclude subagent sessions, system prompts, hidden reasoning, and general tool output. When a complete-round boundary makes a summary due, the collector persists a model-free eligibility record and invokes the one-shot semantic dispatcher. The dispatcher leases the explicit job and runs one ephemeral Codex CLI summary worker. The independent five-minute maintenance owner may overlap at most three such model calls, then serializes verified ingestion through the archive locks and records completion, retry, or quarantine. Worker failure does not stop native capture or cancel successful sibling jobs. Python remains available for low-frequency maintenance, retrieval, reconstruction, and summary ingestion.

A malformed rollout is isolated by a content-free fingerprint of its canonical path, byte size, and nanosecond modification time. An unchanged failed source is not reparsed on every event cycle; any source mutation makes it eligible for an automatic retry, and a successful retry removes the derived fault record. The source remains visible as incomplete coverage throughout quarantine. Never delete, rewrite, or replace the rollout to clear this state.

Federation is a separate low-frequency layer and does not change collector ownership of the local archive. By default, imported replicas live in the sibling <archive>-federation-cache, remain read-only, and are omitted from the desktop primary-archive backup. SSH protects and authenticates the transport; the offline .mwxb bundle itself is neither encrypted nor signed. Federation does not use OpenAI login sessions and does not provide automatic public address discovery, NAT traversal, or mobile access. The optional cloud-folder transport uses the user's existing filesystem synchronization client without receiving its account credentials. It signs and encrypts each target-specific envelope before publication and keeps the five-second local collector path unchanged.