ReadonlybudgetFraction of the token budget reserved for episodic (L2) memories.
ReadonlybudgetFraction of the token budget reserved for recent session (L1) turns.
Optional ReadonlyrelevanceMinimum cosine similarity for an L2 memory to be admitted (0-1). Leave it unset to use the
threshold the embedder was calibrated for (Embedder.relevanceThreshold), falling back to
DEFAULT_RELEVANCE_THRESHOLD for an embedder that declares none. Similarity scales differ
by model, so one fixed number cannot suit them all: the lexical LocalEmbedder scores relevant
text near 0.24, while bge-small scores unrelated text near 0.48.
ReadonlymaxMaximum episodic memories to pull from L2 before budget pruning.
ReadonlyrelevanceWeight on cosine similarity in the blended rank score.
ReadonlyrecencyWeight on recency in the blended rank score.
ReadonlyhybridHybrid retrieval: among the threshold-admitted candidates, fuse the embedding (dense) ranking with a BM25 (lexical) ranking and recency via Reciprocal Rank Fusion, so an exact-keyword match the embedder under-ranks still surfaces. Admission (cosine ≥ threshold) is unchanged. Set false for the classic single-channel dense+recency blend.
ReadonlyrerankCross-encoder rerank window: how many of the over-fetched candidates a configured Reranker rescores. Only takes effect when a reranker is passed to the manager.
ReadonlymaxCap on retained turns per session in L1.
ReadonlytokenHow token counts are computed. Defaults to a dependency-free heuristic.
Configuration for a MemoryManager, mirroring the Python
actrone_memory.MemoryConfig. All fields have safe defaults, soMemoryManager.create()needs no arguments for the in-memory on-ramp.