memory-manager

To keep AI project context persistent across sessions, store decisions, architecture, and blockers in a Git-committed JSONL file that agents load on startup. This avoids re-reading the codebase by reusing prior session memory directly.

TF-IDF search with markdown-aware chunking retrieves relevant context by scoring term frequency across stored memory chunks while preserving markdown structure. This keeps notes readable and ensures efficient retrieval of specific decisions or blockers.

Yes, you can use a JSONL-backed store for cross-session memory management. It preserves the history of decisions and architecture through Git commits, allowing AI agents to load prior context during session starts and post-work reviews.

The best way to prune stale memories while keeping high-value decisions is using a value-weighted garbage collector. It evaluates stored context retention value, automatically removing low-priority entries while preserving critical architectural notes for long-term continuity.

You need persistent project memory for AI-assisted development during session starts, task planning, and post-work reviews. It surfaces prior decisions, blockers, and architectural notes, enabling better planning and reducing rework across new sessions.

Yes, cross-session memory management works without external dependencies. It relies on a JSONL-backed store with TF-IDF search and markdown-aware chunking, requiring no additional libraries to ingest and query project memories.