How memory systems work
Persistent external memory for models and agents.
Our primary subjects are external memory systems that retain state across interactions. Model weights and larger context windows alone are not this product category. Long-context and no-memory configurations remain separately labeled baselines. Injecting retrieved memory into context does not make an external memory system out of scope.
Read the core definition and eligibility criteria.
Reviewed against the linked primary sources on 2026-10-04. These are architecture summaries, not rankings or evidence of benchmark admission. Products change; each result must identify its exact version and settings.
As a simple example, remembering a changed delivery address requires more than finding an old message: the system must select the current fact and preserve enough evidence to justify it. A graph, a memory block and a searchable document offer different ways to organize that information.
GBrain
GBrain adds explicit persistent memory to existing agents. Its documentation describes sourced facts, corrections, withdrawal and shared access across agents, with keyword retrieval and optional semantic search and synthesis. These are documented capabilities, not verified Mnemetric results.
MemPalace
MemPalace is an external memory system for AI interactions. Its project organizes stored material using a memory-palace structure and offers local storage and retrieval. Its benchmark claims require protocol and metric review before comparison; inclusion here does not endorse a ranking.
Mem0
Mem0 extracts useful facts from interactions and retrieves them for later conversations. Its current project description combines semantic similarity, keyword matching and entity links. The managed service includes proprietary optimizations, so an open-source SDK run and a hosted-service run must be identified separately.
Graphiti and Zep
Graphiti represents people, things and their relationships in a graph that tracks when facts are valid and where they came from. Retrieval combines meaning, keywords and graph connections. Zep provides managed context infrastructure built around this approach; it is a distinct deployment from the open-source Graphiti framework.
Letta
Letta gives an agent persistent state. Named memory blocks hold editable text; attached blocks remain in the agent's context, and blocks can be shared between agents. Tools let the agent update that memory. Evaluating the agent therefore requires disclosing its model, tools and memory setup, not only a document search configuration.
Cognee
Cognee turns source material into searchable chunks, entities and relationships. Its retrieval can select graph, vector or code context. Session learning can feed accepted lessons into longer-term memory. The extraction model, embedding model, storage backend and retrieval strategy are part of the evaluated configuration.
MemOS
MemOS organizes memory through operations for storing, retrieving, editing and deleting it. Memory cubes group knowledge for controlled sharing and composition, and its documented inputs include text, images and tool traces. Hosted, self-hosted and local-plugin deployments differ; a result must name the one actually measured.
Supermemory
Supermemory describes a learning component that decides what to retain and relate, backed by a temporal graph with vector and full-text search. Applications add material and retrieve relevant memory through its API. This explains the vendor's documented design; it does not independently verify its internal implementation or performance claims.
HippoRAG
HippoRAG combines document retrieval with a knowledge graph and Personalized PageRank, which spreads relevance through connected nodes. Its interface distinguishes retrieving passages from answering with an LLM. A retrieval score and the quality of the resulting answer measure different stages, even when they use the same index.
Mnemosyne — operator entry
Mnemosyne keeps content-addressed evidence and derives searchable projections from it. It combines hybrid retrieval, time-aware beliefs and provenance, with local SQLite and PostgreSQL deployment paths. Its own documentation records unfinished capabilities and validation gaps. This site's operator must provide the same measured evidence required from every other entry.
Architecture does not determine the winner
Extraction may lose details, retrieval may miss relevant evidence, and an answering model may misread what it receives. Measure these stages separately, together with latency, resource use and safety under the registered protocol.