Skip to content

JiuwenSwarm is an intelligent AI Agent built on openJiuwen. It extends the powerful capabilities of large language models directly to your fingertips through various communication apps you use daily.

Python
5.0 KB
Raw Blame

Coding Memory

Coding Memory is a dedicated memory system for Code mode that allows the Agent to read, write, and edit persistent code context during coding tasks.


Enabling

Coding Memory is automatically enabled when switching to Code mode — no additional configuration required. Related settings are in the modes.code section of config/config.yaml:

modes:
  code:
    rails:
      - FileSystemRail
      - SkillUseRail
      - LspRail
    tools:
      - web_free_search
      - web_fetch_webpage
      - web_paid_search
      - user_todos
    embedding_config:
      model_name: null          # Empty = use global embedding config
      base_url: null
      api_key: null

Tools

Coding Memory provides three tools, automatically registered by CodingMemoryRail within the subprocess:

coding_memory_read

Read current coding memory content.

Parameter Type Description
query string Search keyword to retrieve relevant coding context

Returns matching code snippets and context descriptions.

coding_memory_write

Write new code context or experience to coding memory.

Parameter Type Description
content string Code content or experience to write
section string Category label (e.g. architecture, bugfix, pattern)

coding_memory_edit

Edit an existing coding memory entry.

Parameter Type Description
memory_id string ID of the memory entry to edit
content string New content
section string Optional; change the category label

Relationship to Other Memory Systems

Memory System Applicable Modes Description
Built-in memory (write_memory / read_memory) Agent mode General conversation memory
Experience memory (experience_retrieve / experience_learn) All modes Task-level experience retrieval and consolidation
Coding memory (coding_memory_read / write / edit) Code mode Code context and coding experience

All three memory systems can coexist in Code mode.


Storage

Coding memory is isolated from built-in memory and stored per project under the Agent data workspace:

<agent_workspace>/coding_memory/<project_name>/

When no project is bound, project_name is default. MEMORY.md is the memory index, other top-level Markdown files contain the memories, and files such as memory.db are rebuildable search indexes.

Upgrading from the legacy backend

The legacy backend stored project coding memory directly in <project_dir>/coding_memory/. The first time the project enters Code mode, or an automatic coding-memory extraction runs, its top-level Markdown files are merged into the new project bucket automatically:

  • The legacy directory remains unchanged for rollback.
  • Identical content is deduplicated by SHA-256. A same-name file with different content is imported as <stem>__legacy_<hash8>.md; new data is never overwritten.
  • Databases, lock files, caches, subdirectories, and symbolic links are not migrated. The new backend rebuilds its search index from Markdown.
  • Migration is idempotent. If a rollback adds legacy memories, the changed source fingerprint causes a later initialization to merge only the new content.
  • The target records .coding-memory-migration-v1.json. A failed attempt does not block Code mode or mark that source complete, so the next initialization retries it.
  • Run the first migration after stopping writes from the legacy backend, and avoid concurrent MEMORY.md writes from other sessions. The Rail index lock is process-local and cannot coordinate with a cross-process upgrade migration. Memory content remains available to full-text indexing if such a race drops an index link, but the link may need to be rebuilt.

Only an Agent workspace that can be identified unambiguously as the default space considers the legacy service_default/agent_default and flat legacy workspace paths. A non-default workspace_key does not import them even when no project is bound. The system does not scan arbitrary service_id/agent_id directories because their ownership cannot be inferred safely.


Sleep-Time Consolidation (Dreaming)

When Dreaming is enabled in Code mode, the system reviews past coding sessions during idle time and automatically extracts reusable experience — debugging root causes, API edge behaviors, design decisions — to {workspace}/coding_memory/consolidated_{hash}.md. This is the sleep-time complement to the in-session coding_memory_* tools above.

See Memory → Dreaming: Sleep-Time Memory Consolidation for how to enable it and extraction details.


See Also

  • Modes — Mode configuration and switching
  • Memory — Built-in memory and external memory
  • Task Memory — Task-level experience retrieval