Memory notes

Current anyCode behavior

• Backends: memory.backend supports file, hybrid, noop, or pipeline (归根通道: ephemeral buffer → reinforce → hot Sled → optional vector; see anycode_memory::RootReturnMemoryPipeline). Aliases: layered, guigen.
• Legacy Markdown: With pipeline and memory.pipeline.merge_legacy_file_recall (default true), existing *.md under the memory root are merged read-only into recall alongside the hot layer.
• Scope: Project vs user memories flow through anycode_memory with keyword retrieval today; pipeline adds pre-semantic fragments before promotion.
• Autosave: Controlled by memory.auto_save and successful task completion hooks in the agent runtime. With pipeline, autosave ingests into the buffer (not a direct durable write); repeated touches promote to the hot store.

Optional memory.pipeline JSON fields include: buffer_ttl_secs, max_buffer_fragments, promote_touch_threshold, reinforce_on_recall_match, merge_legacy_file_recall, buffer_wal_enabled, buffer_wal_fsync_every_n, hook_after_tool_result, hook_after_agent_turn, hook_max_bytes, hook_tool_deny_prefixes, embedding_enabled, embedding_model, embedding_base_url, embedding_provider, embedding_local_cache_dir.

• WAL: With buffer_wal_enabled (default true), the ephemeral buffer is appended to a *.pipeline.buffer.wal JSONL file next to the hot Sled DB and replayed on startup. The last fsync happens periodically (buffer_wal_fsync_every_n), after each unit of work in long-running bridges (Telegram, Discord, built-in scheduler, WeChat task, run/orchestration-triggered work), and again when the pipeline store is dropped (normal process exit).
• Vectors: With embedding_enabled or a non-empty embedding_model, or embedding_provider set to local, the pipeline stores vectors in *.pipeline.vec.sled (cosine retrieval).
  • embedding_provider: http (default) uses OpenAI-compatible POST …/embeddings and llm.api_key (override host with embedding_base_url).
  • local: ONNX Runtime via FastEmbed (all-MiniLM-L6-v2, downloaded on first use). Requires building anycode with --features embedding-local. Optional embedding_local_cache_dir overrides the model cache directory (otherwise fastembed’s default, often under ~/.cache/fastembed).
• CLI import: anycode memory import [--dry-run] [--limit N] copies legacy Markdown memories from memory.path into the pipeline hot store (requires memory.backend: pipeline).

OpenClaw parity (research backlog)

OpenClaw ships memory as an extension with its own retention and recall policies. A practical parity checklist:

1. Write triggers: when notes are created (task success only vs tool-driven vs explicit commands).
2. Retrieval: keyword vs embedding / hybrid; per-project isolation guarantees.
3. Compaction interaction: how memories survive /compact and automatic session compression.

Track improvements against this list in issue tracker milestones rather than duplicating OpenClaw internals inside the CLI binary.

Related

• Architecture
• Config & security