ChatDev/docs/user_guide/en/modules/memory.md

Memory Module Guide

This document explains DevAll's memory system: memory list config, built-in store implementations, how agent nodes attach memories, and troubleshooting tips. Core code lives in entity/configs/memory.py and node/agent/memory/*.py.

1. Architecture

1. Memory Store – Declared under memory[] in YAML with name, type, and config. Types are registered via register_memory_store() and point to concrete implementations.
2. Memory Attachment – Referenced inside agent nodes via AgentConfig.memories. Each MemoryAttachmentConfig defines read/write strategy and retrieval stages.
3. MemoryManager – Builds store instances at runtime based on attachments and orchestrates load(), retrieve(), update(), save().
4. Embedding – SimpleMemoryConfig and FileMemoryConfig embed EmbeddingConfig, and EmbeddingFactory instantiates OpenAI or local vector models.

2. Memory Sample Config

memory:
  - name: convo_cache
    type: simple
    config:
      memory_path: WareHouse/shared/simple.json
      embedding:
        provider: openai
        model: text-embedding-3-small
        api_key: ${API_KEY}
  - name: project_docs
    type: file
    config:
      index_path: WareHouse/index/project_docs.json
      file_sources:
        - path: docs/
          file_types: [".md", ".mdx"]
          recursive: true
      embedding:
        provider: openai
        model: text-embedding-3-small

3. Built-in Store Comparison

Type	Path	Highlights	Best for
simple	node/agent/memory/simple_memory.py	Optional disk persistence (JSON) after runs; FAISS + semantic rerank; read/write capable.	Small conversation history, prototypes.
file	node/agent/memory/file_memory.py	Chunks files/dirs into a vector index, read-only, auto rebuilds when files change.	Knowledge bases, doc QA.
blackboard	node/agent/memory/blackboard_memory.py	Lightweight append-only log trimmed by time/count; no vector search.	Broadcast boards, pipeline debugging.

All stores register through register_memory_store() so summaries show up in UI via MemoryStoreConfig.field_specs().

4. MemoryAttachmentConfig Fields

Field	Description
name	Target Memory Store name (must be unique inside stores[]).
retrieve_stage	Optional list limiting retrieval to certain AgentExecFlowStage values (pre, plan, gen, critique, etc.). Empty means all stages.
top_k	Number of items per retrieval (default 3).
similarity_threshold	Minimum similarity cutoff (-1 disables filtering).
read / write	Whether this node can read from / write back to the store.

Agent node example:

nodes:
  - id: answer
    type: agent
    config:
      provider: openai
      model: gpt-4o-mini
      prompt_template: answer_user
      memories:
        - name: convo_cache
          retrieve_stage: ["gen"]
          top_k: 5
          read: true
          write: true
        - name: project_docs
          read: true
          write: false

Execution order:

1. When the node enters gen, MemoryManager iterates attachments.
2. Attachments matching the stage and read=true call retrieve() on their store.
3. Retrieved items are formatted under a "===== Related Memories =====" block in the agent context.
4. After completion, attachments with write=true call update() and optionally save().

5. Store Details

All memory stores persist a unified MemoryItem structure containing:

• content_summary – trimmed text used for embedding search.
• input_snapshot / output_snapshot – serialized message blocks (with base64 attachments) preserving multimodal context.
• metadata – store-specific telemetry (role, previews, attachment IDs, etc.). This schema lets multimodal outputs flow into Memory/Thinking modules without extra plumbing.

5.1 SimpleMemory

• Path – SimpleMemoryConfig.memory_path (or auto). Defaults to in-memory.
• Retrieval – Build a query from the prompt, trim it, embed, query FAISS IndexFlatIP, then apply semantic rerank (Jaccard/LCS).
• Write – update() builds a MemoryContentSnapshot (text + blocks) for both input/output, deduplicates via hashed summary, embeds the summary, and stores the snapshots/attachments metadata.
• Tips – Tune max_content_length, top_k, and similarity_threshold to avoid irrelevant context.

5.2 FileMemory

• Config – Requires at least one file_sources entry (paths, suffix filters, recursion, encoding). index_path is mandatory for incremental updates.
• Indexing – Scan files → chunk (default 500 chars, 50 overlap) → embed → persist JSON with file_metadata.
• Retrieval – Uses FAISS cosine similarity. Read-only; update() unsupported.
• Maintenance – load() checks file hashes and rebuilds if needed. Store index_path on persistent storage.

5.3 BlackboardMemory

• Config – memory_path (or auto) plus max_items. Creates the file in the session directory if missing.
• Retrieval – Returns the latest top_k entries ordered by time.
• Write – update() appends the latest snapshot (input/output blocks, attachments, previews). No embeddings are generated, so retrieval is purely recency-based.

6. EmbeddingConfig Notes

• Fields: provider, model, api_key, base_url, params.
• provider=openai uses the official client; override base_url for compatibility layers.
• params can include use_chunking, chunk_strategy, max_length, etc.
• provider=local expects params.model_path and depends on sentence-transformers.

7. Troubleshooting & Best Practices

• Duplicate names – The memory list enforces unique memory[] names. Duplicates raise ConfigError.
• Missing embeddings – SimpleMemory without embeddings downgrades to append-only; FileMemory errors out. Provide an embedding config whenever semantic search is required.
• Permissions – Ensure directories for memory_path/index_path are writable. Mount volumes when running inside containers.
• Performance – Pre-build large FileMemory indexes offline, use retrieve_stage to limit retrieval frequency, and tune top_k/similarity_threshold to balance recall vs. token cost.

8. Extending Memory

1. Implement a Config + Store (subclass MemoryBase).
2. Register via register_memory_store("my_store", config_cls=..., factory=..., summary="...") in node/agent/memory/registry.py.
3. Add FIELD_SPECS, then run python -m tools.export_design_template ... so the frontend picks up the enum.
4. Update this guide or ship a README detailing configuration knobs and boundaries.