New Unified Memory System

[joaomdmoura]

Summary

Replaces the four separate memory types (ShortTermMemory, LongTermMemory, EntityMemory, ExternalMemory) with a single Memory class where both encoding and retrieval are LLM-driven. Results are ranked by a composite score blending semantic similarity, recency, and importance -- not just vector distance.

Key changes

Unified Memory class.

One API: remember(), recall(), forget(). On save, the LLM infers scope, categories, and importance. On recall, a multi-step RecallFlow assesses query complexity, monitors its own confidence, and deepens search when uncertain. A ConsolidationFlow on save detects near-duplicate records and merges/updates/deletes them automatically.

Composite scoring.

semantic_weight * similarity + recency_weight * decay + importance_weight * importance. Applied in both shallow and deep recall paths, including the confidence router that decides "synthesize vs explore deeper."

Hierarchical scopes.

Memories organize into a tree (/company/engineering, /project/alpha). MemoryScope restricts operations to a subtree; MemorySlice combines multiple scopes into one view.

Config lives on Memory.

All tunable knobs are kwargs: Memory(recency_weight=0.5, exploration_budget=2). No separate config object needed.

Simplified Crew.memory.

Single field replaces memory + memory_instance. Crew(memory=True) creates defaults; Crew(memory=Memory(...)) for custom config.

##$ Memory TUI.
crewai memory opens a terminal UI for browsing scopes, viewing entries with pagination, and running recall queries. textual is now a core dependency.

##$ Flow proxy fix.
LockedListProxy and LockedDictProxy now subclass list/dict so isinstance checks pass in LanceDB, Pydantic, etc.

How memory integrates across the stack

Crew.

Crew(memory=True) creates a default Memory(). Pass Crew(memory=Memory(...)) for custom tuning. All agents in the crew share it. After each task, the executor calls extract_memories() to decompose the raw result into atomic facts, then remember() on each -- so the crew's knowledge base grows automatically as tasks complete.

Agent.

• During task execution, the agent recalls relevant memories and injects them into the task context. The agent uses the crew's shared memory unless it has its own memory field set, in which case it uses that (can be a Memory, MemoryScope, or MemorySlice).

LiteAgent.

Same pattern: LiteAgent(memory=True) or LiteAgent(memory=Memory(...)). Before each turn, it recalls memories matching the user's latest message and appends them to the system prompt. After each turn, it extracts and remembers facts from the response.

Flow.

Every Flow automatically gets a Memory instance. Flow methods can call self.remember(), self.recall(), and self.extract_memories() directly -- memory is a first-class capability, not something wired up externally.

What was removed

ShortTermMemory, LongTermMemory, EntityMemory, ExternalMemory, ContextualMemory, and their storage backends (rag_storage.py, ltm_sqlite_storage.py, mem0_storage.py). All replaced by unified Memory + LanceDBStorage.