Skip to main content
Long-term memory is stored in workspace files, persisting across sessions. The Agent loads historical memory on demand via retrieval tools during conversation, and automatically writes conversation summaries to long-term memory when context is trimmed. Memory Architecture

​
Memory Types

​
Core Memory (MEMORY.md)

Stored in ~/cow/MEMORY.md, containing long-term user preferences, important decisions, key facts, and other information that doesn’t fade over time. The Agent reads and writes this file via tools to maintain long-term knowledge.

​
Daily Memory (memory/YYYY-MM-DD.md)

Stored in ~/cow/memory/ directory, named by date (e.g., 2026-03-08.md), recording daily conversation summaries and key events. Files are only created on first write to avoid generating empty files.

​
Dream Diary (memory/dreams/YYYY-MM-DD.md)

A byproduct of the Deep Dream (memory distillation) process, recording discoveries, deduplication operations, and new insights from each consolidation. Stored in ~/cow/memory/dreams/ directory, named by date.

​
Automatic Writing

The Agent automatically persists conversation content to long-term memory through the following mechanisms:
  • On context trimming — When conversation turns or tokens exceed the configured limit, the oldest half of the context is trimmed, and the discarded content is summarized by LLM into key information and written to the daily memory file. The summary is also asynchronously injected into the retained context for conversational continuity
  • Daily scheduled summary — A full summary is automatically triggered at 23:55 every day, ensuring memory is preserved even on low-activity days (skipped if content hasn’t changed)
  • Deep Dream (memory distillation) — Runs automatically after the daily summary, distilling daily memories into MEMORY.md and generating a dream diary
  • On API context overflow — When the model API returns a context overflow error, the current conversation summary is saved as an emergency measure
All memory writes run asynchronously in a background thread (LLM summarization + file writing), never blocking normal conversation replies.

​
Memory Retrieval

The memory system supports hybrid retrieval modes:
  • Keyword retrieval — FTS5 full-text index matching with BM25 ranking
  • Vector retrieval — Embedding-based semantic similarity search, finds relevant memory even with different wording
The Agent automatically triggers memory retrieval during conversation as needed, incorporating relevant historical information into context. Results are ranked by rank-normalizing the vector and keyword scores onto a common scale and taking a weighted average (default: 0.7 vector weight + 0.3 keyword weight). Because the top hit is always normalized to 1.0, min_score acts as a rank-position cutoff rather than an absolute relevance floor. Daily memory scores decay over time (30-day half-life), while core memory does not decay.
For more precise ranking, rerank can optionally be enabled: run pip install sentence-transformers, then set "rerank_provider": "local" in the config, and results are reordered by a local rerank model (default BAAI/bge-reranker-base). It is off by default and has no effect unless configured.
Files related to memory in the workspace (default ~/cow):

​
Web Console

The memory management page in the Web console allows browsing memory files and dream diaries, with tab switching support:

​
Configuration