mem-orchestrator（内存协调器）

Memory Orchestrator

Implement a layered memory system that remembers the right things, loads the least necessary context, and improves over time through reflection.

Core Model

Represent memory in five layers:

1. Session state — current task, active domains, active objects, constraints
2. Daily raw log — important interactions and memory-worthy events from the day
3. Topic index — summary cards for broad domains. Seed topics like technology/career/investing/research/life/meta are only defaults, not a fixed ontology.
4. Object store — durable objects such as papers, concepts, frameworks, decisions, preferences, open questions, notes, people, projects, or any future type that proves useful
5. Reflection output — compressed updates that merge duplicates, raise abstraction, and add cross-topic links

Do not use one giant memory file as the primary operating model.

Default Directory Layout

Use this layout inside the workspace unless the user specifies another storage backend. Treat it as a readable default, not a rigid schema:

memory/
  session-state.yaml
  working-buffer.md
  daily/
    YYYY-MM-DD.md
  topics/
    technology.yaml
    career.yaml
    investing.yaml
    research.yaml
    life.yaml
    meta.yaml
  objects/
    papers/
    concepts/
    frameworks/
    decisions/
    preferences/
    open-questions/
  reflections/
    YYYY-MM-DD.md
  indexes/
    manifest.json

Cost Control Model

Do not run the full memory pipeline on every message.

Use three levels:

1. Gate only — cheap check for whether the message is memory-relevant at all
2. Conditional recall/write — only when the gate says the message is likely to benefit from memory
3. Low-frequency reflection — background/manual maintenance, never every turn

Use scripts/should_trigger_memory.py or scripts/memory_cli.py turn ... for the default low-cost path.

Workflow

1. Gate the incoming message

Skip memory work for lightweight messages like acknowledgements, simple confirmations, or operational noise.

2. Classify the incoming message

Estimate:

• intent: ask | compare | decide | reflect | write-report | update-preference
• domains: one or more of technology/career/investing/research/life/meta
• whether memory recall is needed
• whether the message contains a memory-worthy event

Use scripts/classify_memory_input.py for a deterministic rule-based baseline.

2. Capture write-ahead events before responding

Before composing the answer, extract memory-worthy events such as:

• explicit preferences
• corrections
• decisions
• stable background facts
• new high-value objects
• relations between known objects

Write them immediately to:

• memory/session-state.yaml
• memory/daily/YYYY-MM-DD.md

Use scripts/extract_memory_events.py and scripts/apply_memory_events.py.

3. Recall with progressive disclosure

Do not load full memory by default.

Recall in this order:

1. Topic cards only
2. Matching object summaries only
3. Full object details only if needed
4. Adjacent associated objects only if they improve the answer

Use scripts/recall_memory.py.

Default retrieval budget:

• topics: top 1-2
• object summaries: top 3-5
• full detail expansions: top 1-2
• associative expansions: top 0-2

4. Answer using the smallest sufficient context

Prefer answers grounded in:

• active topic summary
• the best matching object summaries
• any critical stable preferences

Do not stuff the answer prompt with raw daily logs unless the user is explicitly asking for chronology.

5. Reflect and compress

Periodically review recent daily logs and update the durable layers.

Reflection should:

• merge duplicate objects
• promote repeated facts into preferences or decisions
• update topic summaries
• infer relations like similar_to, contrasts_with, applies_to, extends
• lower the weight of stale or one-off material

Use scripts/reflect_memory.py.

Extensibility Rules

• Do not assume the default topics are exhaustive.
• Add a new topic card when a recurring theme cannot be cleanly routed through existing topics.
• Do not assume the default object types are exhaustive.
• Add a new object subtype directory when repeated objects share a stable shape and retrieval value.
• Prefer human-readable expansion over hidden abstraction.
• When extending the ontology, also update the memory index and README files so the user can inspect the new structure.

Data Shapes

Topic card

id: research
name: 科研 / 论文 / 方法论
summary: 用户会围绕论文研读、研究方法、观点对比进行持续讨论。
subtopics:
  - llm
  - evaluation
recent_objects:
  - paper/constitutional-ai
linked_topics:
  - technology
  - career
stable_preferences:
  - 先讲核心问题，再展开细节
priority_rules:
  - 当问题涉及论文、方法、实验时优先激活

Object

id: paper/constitutional-ai
type: paper
domain: research
title: Constitutional AI
summary: 一篇关于用 AI feedback 替代部分人工 feedback 的论文。
why_it_matters: 对齐训练与偏好学习的重要参照。
tags:
  - alignment
  - preference-learning
status: discussed
confidence: high
last_discussed: 2026-03-30
relations:
  similar_to:
    - paper/rlhf-overview
user_takeaways:
  - 用户更关心训练范式差异而不是宣传式总结

Session state

session_id: auto
active_domains:
  - research
active_objects:
  - paper/constitutional-ai
current_goal: 比较方法差异
recent_constraints:
  - 当前回答要技术向
last_updated: 2026-03-30T01:00:00+08:00

Retrieval Rules

Use a mixed ranking score:

final_score =
  0.35 * semantic_similarity
+ 0.20 * domain_match
+ 0.15 * recency_score
+ 0.15 * relation_score
+ 0.10 * user_preference_match
+ 0.05 * stability_score

If no semantic model is available, approximate with:

• keyword overlap
• explicit domain routing
• recency decay
• relation count
• preference/tag overlap

Engineering Rules

• Write first, answer second, when the user reveals durable information.
• Keep topic cards short; they are routing objects, not full notebooks.
• Keep object summaries decisive enough for summary-first retrieval.
• Do not promote every casual message into durable memory.
• Prefer merging and restructuring over endless append-only growth.
• Treat daily logs as raw material, not as the primary retrieval layer.
• Use associative recall only when it materially improves the answer.

When to create or update objects

Create or update an object when one of these is true:

• the user returns to the same concept across multiple conversations
• the item will likely matter again for future reasoning
• the user asks for comparison, planning, or decision support around it
• the object captures a persistent preference, framework, paper, or decision

Do not create durable objects for:

• one-off factual lookups with no reuse value
• shallow chit-chat
• ephemeral administrative noise

Scripts

scripts/should_trigger_memory.py

Run a very cheap gate to decide whether the current message should trigger memory work at all.

scripts/classify_memory_input.py

Classify raw user text into intent/domains and whether recall/write is likely needed.

scripts/extract_memory_events.py

Extract structured memory events from a single message or turn transcript.

scripts/apply_memory_events.py

Apply events to session state and daily log, and optionally update topics/objects.

scripts/recall_memory.py

Perform summary-first retrieval across topics and objects, returning a compact JSON payload.

scripts/reflect_memory.py

Read recent daily logs and produce compressed updates for topics, objects, and relations.

scripts/memory_cli.py

Provide a unified entry point for bootstrap, gate, turn, capture, recall, reflect, and creating new topics/objects.

References

Read these references when implementing or extending the system:

• references/file-layout.md for storage conventions and field semantics
• references/retrieval-strategy.md for ranking, recall budgets, and expansion rules
• references/reflection.md for compression and promotion rules
• references/openclaw-integration.md for how to wire this into OpenClaw workflows
• references/object-models.md for current topic/object/session schemas and extension rules

For broader maintainability, also read:

• ARCHITECTURE.md
• ROADMAP.md
• EXAMPLES.md