Choose the memory boundary that matches the way your agents actually work.
This is not a pretend winner table. It maps the documented product boundary of each system to the job it is designed to do, using the linked primary source for every competitor.
Mem0
Multi-level memory, hybrid retrieval, entity linking, and temporal reasoning.
Choose it when: Teams choosing an SDK or managed memory platform.
Read primary source →Zep / Graphiti
Temporal knowledge graphs, governed context assembly, and enterprise context infrastructure.
Choose it when: Teams building graph-centric, service-operated agent context.
Read primary source →Letta
In-context memory blocks, archival memory, files, and agent-managed retrieval.
Choose it when: Builders who want to operate agents inside a dedicated runtime.
Read primary source →LangMem
Hot-path tools, background extraction, prompt refinement, and LangGraph storage integration.
Choose it when: Teams already standardized on LangGraph.
Read primary source →Supermemory
Memory, profiles, RAG, connectors, and personal-assistant workflows.
Choose it when: Teams seeking an API-led context stack or personal-memory app.
Read primary source →Memobase
Profile construction, buffered processing, and per-user event memory.
Choose it when: Product teams optimizing personalized user experiences.
Read primary source →Where SuperLocalMemory fits
Choose SLM when your developers need a local-first operating control plane, not only an API or agent runtime: persistent dated evidence, profile-isolated workspaces with personal/shared/global memory boundaries, graph-aware retrieval, cache and compression controls, trusted-peer Mesh, and the same system exposed through MCP, CLI/JSON, hooks, dashboard, and IDE integrations.
Mode A keeps the core memory path local after required assets are present. Modes B and C add local-LLM or external-provider capabilities by explicit choice; optional connectors and Mesh have separate network boundaries.
Published V3 LoCoMo evidence is carried into V3.7 with its original scopes and answer-construction paths disclosed. This page does not rank unlike protocols.
MODE A RAW
60.4%
10 conversations / 1,276 questions. Local embeddings, local retrieval, zero-LLM answer construction.
MODE A RETRIEVAL
74.8%
10 conversations / 1,276 questions. Local retrieval with GPT-4.1-mini answer synthesis.
MODE C
87.7%
Conv-30 / 81 questions. text-embedding-3-large plus GPT-4.1-mini generation and judge.
Published V3 architecture evidence carried into V3.7. These are original protocol-scoped results, not a fresh V3.7 rerun or an ordinal comparison with competitor claims.
Dated memory, multi-channel retrieval, graph evidence, cache, compression, and trusted-peer coordination are operated from one runtime rather than assembled as separate products.
Three public preprints, protocol-scoped LoCoMo evidence, CLI JSON traces, and release-linked tests let developers inspect the mechanism rather than accept a leaderboard slogan.
The same local system is available through MCP, structured CLI, hooks, dashboard, skills, and supported IDE configurations—without rewriting a user’s entire tool configuration.
