Smart Memory v3.1 Skill
Smart Memory v3.1 is a local transcript-first cognitive memory runtime with revision-aware derivation, pinned context lanes, entity-aware retrieval, and bounded prompt composition.
Core runtime:
- - Node adapter: INLINECODE0
- Local API: INLINECODE1
- System facade: INLINECODE2
- Canonical store:
storage/sqlite_memory_store.py plus INLINECODE4
Core Capabilities
- - transcript-first ingest and per-message transcript logging
- typed long-term memory including
preference, identity, and INLINECODE7 - evidence-backed revision lifecycle decisions and supersession chains
- explicit core and working memory lanes
- entity-aware retrieval with lightweight relationship hints
- deterministic rebuild from transcript history
- hot-memory compatibility projection for working context
- strict token-bounded prompt composition with trace metadata
- inspection endpoints for transcripts, evidence, history, lanes, and eval runs
OpenClaw Integration
Use the native wrapper package in skills/smart-memory-openclaw/.
Primary exports:
- - INLINECODE9
- INLINECODE10
The wrapper remains stable while the backend is now transcript-first under the hood.
Tool Interface
- 1. INLINECODE11
- - purpose: query relevant memory through INLINECODE12
- supports
query, type, limit, min_relevance, and optional INLINECODE17 - health-checks the backend before execution
- 2. INLINECODE18
- - purpose: persist important facts, decisions, beliefs, goals, or session summaries
- health-checks the backend before execution
- serializes commits to protect local embedding throughput
- queues failed commits in INLINECODE19
- 3. INLINECODE20
- - purpose: surface pending background insights
- health-checks the backend before execution
- calls INLINECODE21
API Endpoints
Core endpoints:
- - INLINECODE22
- INLINECODE23
- INLINECODE24
- INLINECODE25
- INLINECODE26
- INLINECODE27
- INLINECODE28
- INLINECODE29
Transcript and inspection endpoints:
- - INLINECODE30
- INLINECODE31
- INLINECODE32
- INLINECODE33
- INLINECODE34
- INLINECODE35
- INLINECODE36
- INLINECODE37
- INLINECODE38
- INLINECODE39
- INLINECODE40
- INLINECODE41
- INLINECODE42
- INLINECODE43
- INLINECODE44
Operating guidance
- - query memory before speaking when continuity matters
- do not claim prior context unless retrieval actually supports it
- transcripts are canonical, memories are derived
- treat SQLite as canonical runtime storage
- treat JSON as offline export or backup only
- keep CPU-only PyTorch policy intact
Deprecated
Legacy vector-memory CLI artifacts remain deprecated and should not be revived.
智能记忆 v3.1 技能
智能记忆 v3.1 是一个本地转录优先的认知记忆运行时，具备版本感知推导、固定上下文通道、实体感知检索和边界提示组合功能。
核心运行时：
- - 节点适配器：smart-memory/index.js
- 本地 API：server.py
- 系统外观：cognitivememorysystem.py
- 规范存储：storage/sqlitememorystore.py 加 transcripts/
核心能力
- - 转录优先摄取和每条消息的转录日志记录
- 类型化长期记忆，包括 preference、identity 和 task_state
- 基于证据的修订生命周期决策和替代链
- 显式的核心和工作记忆通道
- 带有轻量级关系提示的实体感知检索
- 基于转录历史的可确定性重建
- 工作上下文的热记忆兼容性投影
- 带有跟踪元数据的严格令牌边界提示组合
- 转录、证据、历史、通道和评估运行的检查端点
OpenClaw 集成
使用 skills/smart-memory-openclaw/ 中的原生包装包。
主要导出：
- - createSmartMemorySkill(options)
- createOpenClawHooks({ skill, agentIdentity, summarizeWithLLM })
包装器保持稳定，而后端现在底层采用转录优先架构。
工具接口
- 1. memory_search
- - 目的：通过 /retrieve 查询相关记忆
- 支持 query、type、limit、minrelevance 和可选的 conversationhistory
- 执行前对后端进行健康检查
- 2. memory_commit
- - 目的：持久化重要事实、决策、信念、目标或会话摘要
- 执行前对后端进行健康检查
- 序列化提交以保护本地嵌入吞吐量
- 将失败的提交排队到 .memoryretryqueue.json
- 3. memory_insights
- - 目的：展示待处理的背景洞察
- 执行前对后端进行健康检查
- 调用 /insights/pending
API 端点
核心端点：
- - GET /health
- POST /ingest
- POST /retrieve
- POST /compose
- POST /runbackground
- GET /memories
- GET /memory/{memoryid}
- GET /insights/pending
转录和检查端点：
- - POST /transcripts/message
- GET /transcripts/{sessionid}
- GET /transcript/message/{messageid}
- GET /memory/{memoryid}/evidence
- POST /revise
- GET /memory/{memoryid}/history
- GET /memory/{memoryid}/active
- GET /memory/{memoryid}/chain
- GET /lanes/{lanename}
- POST /lanes/{lanename}/{memoryid}
- DELETE /lanes/{lanename}/{memoryid}
- POST /rebuild
- POST /rebuild/{sessionid}
- GET /eval/suite/{suitename}
- GET /eval/case/{caseid}
操作指南
- - 在需要连续性的对话前查询记忆
- 除非检索实际支持，否则不要声称拥有先前的上下文
- 转录是规范的，记忆是推导的
- 将 SQLite 视为规范的运行时存储
- 仅将 JSON 视为离线导出或备份
- 保持纯 CPU PyTorch 策略不变
已弃用
遗留的向量记忆 CLI 工件仍处于弃用状态，不应恢复。
