Architectures

How memory systems work.

Eight ways to remember things and find them again. Read from each project’s own docs on 2026-10-04.

1 · CaptureWhat gets kept from a conversation.
2 · OrganiseHow it is stored and updated.
3 · RetrieveWhat comes back for a question.
4 · AnswerWhat the model does with it.

A system can be strong at one stage and weak at the next, which is why a single score never tells you much.

Mem0

Fact extraction

VectorKeywordEntity linksHosted or OSS

Pulls facts out of conversations and finds them again later, using meaning, keywords and entity links.

Watch for: The paid service has extras the open-source package does not. Scores must say which one ran.

Primary source

Graphiti and Zep

Temporal knowledge graph

GraphValid-time factsProvenanceHybrid search

Stores people, things and relationships as a graph, tracking when each fact was true.

Watch for: Graphiti is the open framework; Zep is the paid service built on it. Different products.

Primary source

Letta

Stateful agents

Memory blocksAgent toolsShared blocks

Gives an agent editable memory blocks that stay in its context and can be shared between agents.

Watch for: The agent's model and tools shape the result as much as the memory does.

Primary source

Cognee

Knowledge pipelines

GraphVectorEntitiesSession lessons

Turns documents into searchable chunks, entities and relationships, then picks graph, vector or code context.

Watch for: The extraction model, embedder and storage backend are all part of what gets measured.

Primary source

MemOS

Memory operating system

Memory cubesMultimodalEdit and delete

Treats memory as an operating system: store, retrieve, edit and delete, grouped into shareable cubes.

Watch for: Hosted, self-hosted and plugin builds differ. A score must name which one ran.

Primary source

Supermemory

Memory API

Temporal graphVectorFull-text

Decides what to keep and how it connects, backed by a time-aware graph with vector and text search.

Watch for: This describes what the vendor documents. We have not verified it.

Primary source

HippoRAG

Graph-assisted retrieval

Knowledge graphPersonalized PageRank

Spreads relevance through a knowledge graph so related documents surface together.

Watch for: Finding passages and answering well are separate stages, even on the same index.

Primary source

Mnemosyne

Evidence-first memory

Operator entry
Evidence ledgerHybrid retrievalTime-aware beliefsProvenanceSQLite / PostgreSQL

Keeps evidence first and builds searchable views from it, with time-aware facts and provenance. Runs on SQLite or PostgreSQL.

Watch for: This is our own system. Its own docs list unfinished parts, and it must show the same evidence as everyone else.

Primary source

What each one can do

Shape, not score: can you run it yourself, does it understand time, can it say where a memory came from.

SystemOpen sourceSelf-hostedGraph memoryTime-aware factsProvenanceMCP serverBest known for
Mnemosyne Operator entryYesYesYesPartialYesYesCalibrated abstention, prospective actions, auditable evidence
Mem0YesYesPartialLimitedLimitedYesSimple API, wide adoption
Zep / GraphitiPartialPartialYesYesPartialYesTemporal knowledge graph
LettaYesYesNoLimitedLimitedYesAgent-managed memory blocks
LangMemYesYesNoLimitedLimitedNoNative fit for LangGraph agents
CogneeYesYesYesPartialPartialYesDocuments to knowledge graph pipelines
MemOSYesYesYesPartialPartialPartialMemory as an operating-system layer
supermemoryPartialPartialPartialLimitedLimitedYesDrop-in memory API and connectors
OpenAI memoryNoNoNoNoNoNoBuilt into ChatGPT, no setup, no control
YesPartialLimited: exists, not a design centreNo

Our own summary, deliberately rough and not verified. Spotted something wrong? Tell us.