@agent-trace-cache/core

Deterministic TypeScript trajectory cache for agentic systems. Caches goal + context → plan rather than query → response (semantic-cache layer) or thread messages (Mastra Memory layer).

The hit path is 100% deterministic TypeScript — no LLM is invoked when the cache returns a result. Embeddings (OpenAI text-embedding-3-small) are precomputed at write time and compared deterministically at read time.

Install

pnpm add @agent-trace-cache/core
# or: npm install @agent-trace-cache/core

For the higher-level createCache / withCache ergonomics, install the SDK package instead:

pnpm add @agent-trace-cache/sdk

What's in this package

• AgentTraceCache — the cache facade with lookup / store / invalidate / recordReplayQuality
• ContextDiff — four normalization pipelines (ephemeral / structural / semantic / discrete) with per-pipeline ablation
• ReplayEngine + TaintRecorder — L1 (full match) / L2 (partial replay via taint analysis) / miss
• HNSWIndex — in-process vector ANN
• Storage adapters — InMemoryTrajectoryStore, LibSqlTrajectoryStore (with opt-in HNSW persistence)
• assertSerializable, hash utilities, embedding helpers

Documentation + benchmarks

• Project overview and quickstart: github.com/kylemaa/agent-trace-cache
• Methodology and τ-bench airline benchmarks (69% hit rate, 90% judge precision, 62.7% L2 partial replay under 4-axis context drift): see blog/agent-trace-cache.md in the repo

License

MIT