agent-memory

To persist context across Claude Code subagent runs, you need a memory system that stores task outcomes as Markdown notes with YAML frontmatter. This allows agents to retrieve past experiences and continuously improve memory quality using feedback learning.

BM25 relevance retrieval for agent memory ranks stored experiences using a three-dimensional scoring system combining recency, importance, and BM25-style text relevance. This ensures the most contextually appropriate memories are retrieved for a given task query.

Yes, you can use YAML frontmatter to manage subagent memory health by categorizing memories as healthy, warning, or blocked. Blocked memories are automatically filtered out from future retrieval and consolidation processes to maintain context quality.

The best way to deduplicate near-duplicate agent memories is using Jaccard similarity merging with threshold-based consolidation. This process combines overlapping context entries and applies Ebbinghaus-inspired exponential decay to reduce stale, low-value memories over time.

Yes, prompt injection works with multi-agent memory stores by enriching an agent prompt with a compact formatted block of top retrieved memories. It supports personal and shared memory stores, optional cross-agent retrieval, and spreading activation via related memory links.

You apply feedback learning to adjust memory usefulness by recording positive or negative feedback on retrieved memories. This feedback-driven scoring adjustment updates the importance dimension, ensuring that memories proven useful are prioritized in future BM25 retrieval operations.