API Reference — Overview

On this page

This section documents every public class exported from the top-level memory module . Classes are grouped by the subpackage that defines them.

Core (`memory.core`)

API	Type	Description
`MemoryType`	`Enum`	Categorizes memories: `SHORT_TERM`, `LONG_TERM`, `WORKING`, `EPISODIC`, `SEMANTIC`, `PROCEDURAL`.
`MemoryPriority`	`Enum`	Priority tiers (`CRITICAL`…`MINIMAL`) derived from an importance score.
`MemoryEntry`	`dataclass`	The universal memory record used by every tier in this package.
`BaseMemory`	`ABC`	Abstract interface implemented by all four `ConversationXMemory` classes.
`BaseRetriever`	`ABC`	Abstract interface for retrieval-style components; not implemented by any concrete class in this source tree.
`MemoryConfig`	`dataclass`	Central configuration object (32 fields — see Configuration).
`ChatMessage`	`dataclass`	Lightweight input/output/timestamp record. Exported but unused internally.
`Entity`	`dataclass`	Named entity with rolling context list, used by `ConversationEntityMemory`.

API	Type	Description
`ConversationBufferMemory`	`class`	Stores the entire conversation, unbounded.
`ConversationBufferWindowMemory`	`class`	Buffer memory capped to the last `k` turns via a `deque`.
`ConversationSummaryMemory`	`class`	LLM-summarizes old turns once a token budget is exceeded.
`ConversationEntityMemory`	`class`	Tracks named entities and their accumulated context.

API	Type	Description
`LongTermMemory`	`class`	SQLite-backed, thread-safe, multi-tenant persistent store.

API	Type	Description
`SensitiveDataMasker`	`class`	Regex-based PII detection & redaction (10 built-in patterns).
`MemoryEncryptor`	`class`	Fernet symmetric encryption with base64 fallback.
`AccessController`	`class`	RBAC for `(user_id, tenant_id, Permission)` triples.
`Permission`	`Enum`	`READ`, `WRITE`, `DELETE`, `ADMIN`.

API	Type	Description
`SemanticMemory`	`class`	Vector similarity store (FAISS or pure-NumPy fallback).

API	Type	Description
`ContextBuilder`	`class`	Assembles documents/memories/history/profile into a token-budgeted prompt.
`Document`	`dataclass`	A single retrieved RAG chunk.
`BuiltContext`	`dataclass`	The result of `ContextBuilder.build()`.

API	Type	Description
`ShortTermMemory`	`class`	Fixed-window `deque` of `MemoryEntry`, decay-aware.
`WorkingMemory`	`class`	Importance-ranked, token-budgeted active context store.
`MemorySelector`	`class`	Multi-factor scoring (relevance/recency/importance) and selection.
`SelectionConfig`	`dataclass`	Weights and thresholds for `MemorySelector`.
`ScoredMemory`	`dataclass`	A candidate entry plus its component and composite scores.
`MemoryCompressor`	`class`	Noise removal, near-duplicate merging, LLM summarization.
`ForgettingMechanism`	`class`	Ebbinghaus-style decay + deletion gating.
`ForgettingConfig`	`dataclass`	Tunables for `ForgettingMechanism`.
`UserProfileManager`	`class`	Per-user profile CRUD with optional JSON persistence.
`UserProfile`	`class`	Typed accessor/mutator wrapper over a profile dict.
`AIMemoryManager`	`class`	Top-level orchestrator tying every subsystem together.

The following are implementation details, not exported from any __init__.py, and therefore not documented as standalone pages:

memory.semantic_memory.semantic_memory._EmbeddingModel, _NumpyIndex, _FAISSIndex
memory.intelligent.memory_manager._TenantBundle
memory.intelligent.memory_compressor.LLMProtocol, memory.intelligent.memory_manager.LLMProtocol (two structurally-identical Protocol definitions duplicated in separate modules rather than shared from one location)

New to the package? See Quickstart.
Need the big-picture wiring between these classes? See Architecture.
Found a discrepancy between a docstring and actual behavior? Check Known Issues & Bugs first — it may already be documented there.