compound-memory
Orchestrate governed persistent project memory across immutable evidence, llm-wiki, line-memory, reviewed session candidates, generated outputs, and bounded context packs.
- License 许可
- MIT
- 所属仓库
- jmagly/aiwg
- 最近更新
- 2026-11-14
- Stars
- ★ 229
- Forks
- 36
01 前言 Front Matter
- name
- compound-memory
- triggers
- [ "use compound project memory", "ingest sources into persistent memory", "review memory candidates", "maintain line memory and wiki", "build a memory context pack", "capture an output into memory" ]
- namespace
- aiwg
- platforms
- [ "all" ]
- description
- Orchestrate governed persistent project memory across immutable evidence, llm-wiki, line-memory, reviewed session candidates, generated outputs, and bounded context packs.
03 SKILL.md 正文
Compound Memory
Use this driver for the persistent compounding-memory pattern. It coordinates existing capabilities; it does not replace their storage, review, indexing, or promotion mechanics.
Start with readiness:
aiwg compound-memory status --json
Governed workflows
- Ingest — preserve the source unchanged, calculate its identity and
digest, then route long-form material through
memory-ingestand llm-wiki. Route session transcripts throughaiwg sessions importandextract. - Retrieve/use — query llm-wiki or the artifact index for linked detail;
use
aiwg compound-memory context <task> --budget <characters> --jsonto merge relevant line facts and wiki evidence. Add--no-touchfor inspection. Never inject the complete stores into startup context. - Write/capture output — register the output and its source/context-pack
lineage before proposing any extracted facts as session candidates. Run
aiwg compound-memory capture-output <file> ... --jsonfirst, review the immutable digest and minimized references, then repeat with--confirm --operation-id <preview-operation-id>. This creates registration and index receipts only; it never promotes knowledge. - Manage — use exact-version candidate review and preview promotion to the
declared
memoryorline-memoryconsumer. One reviewed candidate may produce independent receipts for both consumers. - Review/update — independently accept, reject, or defer proposed facts.
Start from
aiwg compound-memory review --limit 50 --json, then inspect and decide through the exact-version sessions commands. Workspace identity changes require a separate explicit review gate. For canonical context, runaiwg compound-memory update <target> <key> <value> --source-ref <ref> --reviewer <id> --reason <reason> --json, review its diff/conflicts, then repeat with the exact operation ID and--confirm. This never refreshes provider adapters. - Maintain — preview pruning, deduplication, contradiction repair, orphan repair, and index refresh. Confirm only the reviewed mutations. Preserve lifecycle tombstones and source-purge dependent dispositions. The compound command's confirmation may replay pending output registrations; all other actions remain delegated to their owning review workflows.
Until the compound-memory lifecycle ADR is accepted, do not invent a new automatic promotion path. Use the existing commands and receipts named above, and treat generated/model output as untrusted proposed knowledge rather than ground truth.
References
- @$AIWG_ROOT/agentic/code/addons/compound-memory/docs/overview.md
- @$AIWG_ROOT/agentic/code/addons/line-memory/skills/line-memory/SKILL.md
- @$AIWG_ROOT/agentic/code/addons/llm-wiki/skills/llm-wiki/SKILL.md
- @$AIWG_ROOT/agentic/code/addons/semantic-memory/skills/memory-ingest/SKILL.md
04 同仓库更多 Skills
host-hardening
Read-only system hardening baseline and plan-first host hardening recommendations for SSH, Docker bind addresses, firewall persistence, datastores, exporters and audit retention.
exposure-analysis
Read-only firewall analysis and effective host exposure reconstruction for Docker published ports, UFW bypass, stale rules, CDN scope, IPv6 asymmetry and external probe allowlists.
host-hardening
Read-only system hardening baseline and plan-first host hardening recommendations for SSH, Docker bind addresses, firewall persistence, datastores, exporters and audit retention.
exposure-analysis
Read-only firewall analysis and effective host exposure reconstruction for Docker published ports, UFW bypass, stale rules, CDN scope, IPv6 asymmetry and external probe allowlists.
fixture-skill
Exercise rule delivery.
pi-local
Pi-local duplicate-precedence fixture.