Memory forge

• English name: memoryforge
• Chinese name: 记忆锻造

MemoryForge is Cadau’s “conversation insight and knowledge feed” mechanism: by archiving real conversations and extracting frequent and critical questions offline, it keeps improving answer quality without changing human-written knowledge originals.

Method layer: ../索引式文档与反馈闭环.md. This page focuses on how MemoryForge is implemented and how it runs.

Goals

• Let help and team knowledge evolve with real questions, originals read-only, improvements on overlay layers.
• Cut repeat Q&A cost and raise first-answer hit rate.
• Provide stable data input for later “adopt → write overlay / skill / index”.

Originals vs overlay (core principle)

When the agent answers: originals win; overlays fill gaps; on conflict the original wins and the injected block marks the source.

Adopt action types (action_type)

After scan clustering, each proposal carries one of the following (rules + optional LLM polish of suggested copy):

Forbidden: a proposal or one-click apply that rewrites or merges into an existing .md original.

Shipped now

• After chat finishes, the backend writes “user question + assistant reply” to the archive.
• Archive format: daily jsonl files.
• Default directory: tmp/memoryforge/chat-archive
• Config: CHAT_ARCHIVE_DIR
• Scan and proposals (MVP): Admin POST /admin/memoryforge/scan scans the last N days of archive, clusters by repeated questions, writes proposals with action_type to tmp/memoryforge/proposals/; optional LLM polish of top suggestions.
• User rating signal: conversation archive writes assistant_message_id; after the user rates an assistant reply, rating / rating_reasons / rating_note are written back. On scan: help conversations rated “not satisfied” can enter a proposal on a single occurrence, priority high, with suggested copy and evidence.
• Admin page: /memoryforge can scan, view historical proposals and evidence conversations (including “user not satisfied” marks).
• Adopt write (answer supplement): POST /admin/memoryforge/proposals/{id}/adopt writes supplement clusters into _forged/ under the matching knowledge root (new Markdown + index), without changing originals; proposal JSON writes back adopted_path.
• Retrieval merge: help assistant and workspace/agent knowledge retrieval merge _forged/ hit passages and mark “if this conflicts with the original, the original wins”.

Archive records include (example fields):

• timestamp, date
• request_id, session_id
• user_id, workspace_id, user_agent_id
• help_mode, partial
• intent_meta (structured intent: intent, slots, confidence, need_clarification)
• user_message, assistant_message, assistant_message_id
• rating, rating_reasons, rating_note, rated_at (written back after rating)
• topic, review (conversation topic and confirmed review, if any; helps cluster by topic)

Proposal cluster fields (example):

• title, frequency, target_layer (help | workspace)
• action_type (see the table above)
• priority, suggestion, evidence
• dissatisfied_count, rating_reasons (when the user was not satisfied)

Later plan (suggested)

Phased delivery: see ../../记忆锻造与定时任务实施计划.md (MF track).

1. ~~Time-window scan~~: MVP shipped (admin scans last N days).
2. ~~Signal extraction and action classification~~: MVP already has frequency clustering, weak-answer priority, and four action_types.
3. ~~Adopt write (answer supplement)~~: supplement / verified_answer can write _forged/; skill writes the workspace Skills center; index improves index.json aliases.
4. ~~Retrieval merge~~: help and workspace/agent knowledge retrieval already merge _forged/.
5. ~~Light outcome look-back~~: admin last-30-day adopt stats by action_type.
6. Optional auto scan: MEMORYFORGE_AUTO_SCAN_ENABLED environment variable (adopt still human).

Relation to the index loop

• The index loop defines “how to organize documents and how to form an improvement cycle”.
• MemoryForge provides “collect data, analyze offline, propose improvements”.
• Together, the document system can keep evolving:

- the index system owns “retrieval and execution quality”; - MemoryForge owns “find problems and drive overlay updates”, and does not replace humans maintaining originals.

Versus long-term agent memory

OpenClaw Dreaming / memory promotion in the spec maps to the right-hand “agent long-term memory” path in Cadau (shipped: conversation extract, flush before compaction, idle maintain agent_memory_maintain), not MemoryForge. MemoryForge does not promote conversation logs straight into structured memory entries, and does not delete personal memory/entries/. Idle dedupe/expiry: ../智能体记忆.md “Idle maintain”.

Naming

• Mechanism name: MemoryForge
• Proposal directory: tmp/memoryforge/proposals
• Overlay directories (planned): system help/_forged/, workspace knowledge/_forged/

Related mechanisms

• ../索引式文档与反馈闭环.md
• ../智能体对话编排.md
• ../智能体调用知识文档的方式.md