Changelog

All notable changes to Recallm will be documented in this file.

The format is based on Keep a Changelog.

[0.2.0] — 2026-03-10

SemanticCache.stats() returns a CacheStats dataclass with hits, misses, hit_rate, avg_similarity, and namespace_sizes for development-time visibility into cache behaviour
ThreadSafeInMemoryStorage — RLock-protected drop-in replacement for InMemoryStorage, safe for multi-threaded and async-framework use
recallm top-level package: pip install recallm now maps directly to from recallm import ..., eliminating the install/import name mismatch

Default embedding model changed from all-MiniLM-L6-v2 to BAAI/bge-small-en-v1.5 for improved first-run reliability
Default cache_timeout_seconds raised from 0.05 to 0.2 (200 ms) to reduce silent cold-start bypasses
Cache timeout now emits a structlog warning (cache.timeout_exceeded) with elapsed_ms, timeout_ms, and action=bypass instead of a silent error

SemanticCache.wrap() with sync and async support
InMemoryStorage — zero-dependency in-process backend
RedisStorage — persistent backend with lazy tombstone cleanup
FastEmbedEmbedder — ONNX-based, ~20MB default embedder
SentenceTransformerEmbedder — optional torch-based embedder
Three similarity threshold profiles: strict (0.97), balanced (0.92), loose (0.85)
Namespace-based cache invalidation
TTL support on cache entries
Prometheus metrics: hits, misses, errors, embedding latency, similarity scores, stream bypass
Structlog structured logging on all cache events with rich operational fields
stream=True bypass with per-namespace counter
Fail-open on all cache operation failures
CacheContext type alias for type-checking convenience
SemanticCache.async_warmup() for non-blocking model load in async frameworks
Grafana dashboard (8 panels, $namespace variable)
Benchmark suite with four realistic prompt distributions