A lightweight, zero-vector-database, file-system based memory architecture for autonomous multi-agent teams. Features Three-Layer Memory Partitioning (
palace://), Cryptographic Event Sourcing (SHA-256 Hash Chain), Tiered Context Envelopes (L0/L1/L2), and Heartbeat Process Supervision.
Virtual Filesystem: palace://
┌─────────────────────────────────────────────────────────────────────────────┐
│ Layer 1: Invariant Rules (palace://rules/) │
│ • Core safety constraints, governance axioms, immutable review policies │
├─────────────────────────────────────────────────────────────────────────────┤
│ Layer 2: Domain Topology (palace://topology/) │
│ • Workspace asset maps, API boundaries, verification recipes │
├─────────────────────────────────────────────────────────────────────────────┤
│ Layer 3: Event-Sourced Decision Stack (palace://decisions/) │
│ • Append-only decision_events.jsonl (SHA-256 hash chained, monotonic seq) │
│ • Active Stack (stack/), Dead-End Lessons (dead_ends/), History (history/) │
└─────────────────────────────────────────────────────────────────────────────┘
│
┌─────────────────────────┴─────────────────────────┐
▼ ▼
┌───────────────────────────────┐ ┌───────────────────────────────┐
│ Tiered Context Envelopes │ │ Process Supervision │
│ • L0 (~50 tokens, abstract) │ │ • Heartbeat monitoring │
│ • L1 (~300 tokens, context) │ │ • Task stall/hang detection │
│ • L2 (Full JSON/Hash chain) │ │ • Fail-closed state engine │
└───────────────────────────────┘ └───────────────────────────────┘
git clone https://github.com/Keithsys730/mempalace.git
cd mempalace
python -m pip install -e ".[test]"
# Verify installation and run the 132-test offline suite
pytest -q
# Use the registered CLI directly (or via `python -m mempalace.cli`)
mempalace bootstrap
mempalace tree
mempalace context --level L0MemPalace 為多 Agent 協同作業提供 三層式階層記憶 (Three-Layer Memory Palace)、確定性事件溯源回放 (Deterministic Event Replay) 與 進程心跳健康度監督 (Process Supervision)。
- Layer 1: 不可變核心規約 (
palace://rules/):治理不可變常數、安全門禁規則。 - Layer 2: 領域拓撲與資產地圖 (
palace://topology/):各專案檔案結構、API 邊界與驗證配方。 - Layer 3: 確定性事件溯源決策棧 (
palace://decisions/):- 以追加寫入 (Append-Only) 的
decision_events.jsonl為單一事實來源 (Source of Truth)。 - 具備 SHA-256 哈希鏈與單調序列號 (
seq_id),防篡改 (Fail-Closed)。 - 支援
stack/活躍棧、dead_ends/避坑池與history/歷史歸檔。
- 以追加寫入 (Append-Only) 的
- 進程心跳健康度監督 (
palace://supervision/):監督多 Agent 背景任務狀態機與超時假死。 - 事件審計與溯源 (
palace://audit/):防篡改時間序列事件日誌。
L0(Abstract):~30-80 Tokens,單句極簡摘要與統計指標,適合 Prompt 標頭常駐與快篩。L1(Overview):~200-400 Tokens,結構化核心脈絡、規約與活躍決策,適合規劃與換棒。L2(Detail):完整原始 JSON/Schema、完整假設、時間戳與哈希鏈。
安裝後可直接使用 mempalace 命令(或 python -m mempalace.cli):
# 1. 檢視虛擬記憶宮殿目錄樹
mempalace tree
# 2. 巡檢虛擬目錄 (ls) 與檢視節點 (view)
mempalace ls palace://rules --level L0
mempalace view palace://topology/mempalace --level L1
# 3. 全層級遞迴檢索 (find)
mempalace find "mempalace"
# 4. 輸出分級 Context Package
mempalace context --level L0
mempalace context --level L1 --json
# 5. 壓入新決策 (發布 DECISION_PUSHED 事件)
mempalace push-decision -t "Feature A" -v "BUILD" -r "Passed test criteria"
# 6. 彈出決策並保留死胡同 (發布 DECISION_POPPED 事件)
mempalace pop-decision -r "Completed"
# 7. 記錄探索失敗的死胡同路徑 (發布 DEAD_END_RECORDED 事件)
mempalace record-dead-end -p "legacy_endpoint" -r "Deprecated in v2"
# 8. 檢視防篡改事件審計日誌
mempalace audit-log
# 9. 時空穿梭回放至指定事件點
mempalace replay --to-seq 1
# 10. 全量校驗事件鏈防篡改完整性
mempalace verify-integrity
# 11. 檢查背景任務健康度與假死監控 (Process Supervision)
mempalace check-healthThis repository includes a portable SKILL.md. Mount or copy it into ~/.claude/skills/mempalace/, ~/.codex/skills/mempalace/, or ~/.gemini/config/skills/mempalace/ so any AI coding agent can query palace:// memory envelopes and manage event-sourced decisions.