rohitg00/agentmemory

Hybrid Storage Architecture

AgentMemory rejects the naive "just use a vector DB" approach in favor of a tiered persistence model optimized for code semantics:

Design Trade-offs

• Embedding Strategy: Uses code-aware embeddings (CodeBERT-style) rather than general text, sacrificing general document memory for superior function-level recall.
• Synchronization: Async batch writes to prevent I/O blocking on agent inference; risks losing last 30s of context on crash.
• Privacy: Local-only by default—no cloud vector DB—making it enterprise-friendly but limiting cross-device sync without custom backends.

The breakthrough isn't persistence itself—it's the SWE-bench for Memory methodology that validates which context actually improves agent performance on real GitHub issues, not synthetic benchmarks.

Concrete Technical Innovations

• Episodic Memory Clustering: Groups terminal commands, file edits, and LLM reasoning traces into "task episodes" using edit-distance heuristics, enabling retrieval of entire debugging workflows rather than isolated snippets.
• Cross-Session Intent Bridging: Maintains a ProjectGraph that tracks incomplete TODOs across agent restarts, automatically injecting /* You were implementing X but got stuck on Y */ primers into new sessions.
• Agent-Agnostic Protocol: Implements the Model Context Protocol (MCP) to work with Claude Code, Cursor, and Copilot simultaneously without vendor-specific hacks—critical for teams using mixed IDE environments.
• Semantic Diff Compression: Instead of storing full file states, stores AST diffs with natural language annotations, reducing storage footprint by 90% while preserving semantic intent of changes.
• Benchmark-Driven Retrieval: Open-sourced the CodeMemory-Harness evaluation suite testing recall on 500+ real-world coding tasks from GitHub issues, ensuring memory retrieval actually helps agents pass tests rather than just matching embeddings.

Benchmark Results

Scalability Limitations

Current SQLite backend shows O(n) query degradation beyond ~50,000 stored memories (roughly 6 months of heavy coding). The Redis adapter scales horizontally but loses the zero-config advantage. Notably, the compression algorithm struggles with generated code (high entropy, low semantic structure), causing 3x storage spikes when agents write boilerplate-heavy frameworks like React components.

Competitive Landscape

Integration Surface

Currently supports Claude Code (via MCP), Cursor (via .cursorrules hooks), and GitHub Copilot Chat (limited, via VS Code extension API). The project risks platformization risk: if Anthropic builds native persistent memory into Claude Code 2.0, this library becomes unnecessary. However, the cross-agent portability provides vendor-lockin insurance that native solutions cannot match.

Adoption Signals

Heavy traction among AI-assisted OSS maintainers—the 86 forks suggest developers are customizing the memory compression algorithms for specific languages (Rust and Go forks notably popular). Missing enterprise features like team-shared memory or PII scrubbing, keeping it currently in the individual power-user niche.