Mem0: AI Memory Management for LLMs and Agents#

Original | Raw

Technical deep-dive from the mem0 engineering team on production memory management for LLM agents. Provides the most concrete benchmark data in the wiki for memory system tradeoffs.

Four Memory Layers#

1. Conversation memory — active context window. Resets at session end.
2. Session memory — spans a single task/goal sequence. Survives across turns.
3. User memory — long-term: preferences, tools, projects, communication style. Most valuable, most expensive.
4. Organizational memory — team/company level. Shared policies, consistent across all agents.

Each answers a different question: “what’s happening now?” / “what’s this task’s context?” / “what do I know about this person?” / “what’s universally true?”

Two-Phase Extraction Pipeline#

Phase 1 — Extraction: LLM pass identifies discrete, durable facts (not summaries). “User prefers Python,” “user’s timezone is CET.” Mirrors Craik & Lockhart’s levels-of-processing theory.

Phase 2 — Update: Four operations before any write:

• ADD — new fact
• UPDATE — supersedes existing (user changed jobs)
• DELETE — existing fact no longer true
• NOOP — duplicate, skip

Resolves contradictions at write time, not query time. Store stays coherent as it grows. Directly addresses the wiki’s identified tension that “persistent context compounds value AND errors.”

Memory Types (Cognitive Science Mapping)#

• Semantic — factual knowledge (what user does, tools, domain)
• Episodic — event-specific (what happened in a conversation). Decays faster.
• Procedural — how things are done (workflows, patterns). Underused, disproportionately valuable.
• Working — active context window contents

Graph-Enhanced Retrieval (Mem0g)#

Vector similarity weaknesses: phrasing-sensitive, no time concept, no contradiction detection. Mem0g stores memories as directed labeled graph — entities as nodes, relationships as edges. Enables graph traversal for relational queries.

Forgetting as Design Requirement#

Based on Bjork’s “New Theory of Disuse”: forgetting is active and adaptive, protects retrieval quality. Memories decay without reinforcement. Entries below threshold are pruned.

Benchmark Results (LOCOMO)#

93% token reduction with selective retrieval. The 4.5-point accuracy gap vs full-context is the core tradeoff.

Memory Scoping#

Five dimensions: user_id, session_id, agent_id, run_id, org_id. Over-broad scoping bleeds memory between users. Under-broad means agent can’t access what it should.

See Also#

• context-management
• agent-memory-persistence
• continuum-memory-architectures
• agent-memory-systems-2026