If you are building agents in 2026, you have already hit the wall. Bigger models do not fix forgetfulness. Context windows can grow forever, and the agent still cannot remember what a user told it last Tuesday, that the customer's address changed three months ago, or that a recommendation it made last year turned out to be wrong.

This is why "agent memory" has become a category in its own right. Agent memory is distinct from RAG, from vector search, and from the model itself. Below are the ten products I watch most closely, ordered by how I would reach for them when building a real system. I founded MinnsDB, so the ranking is biased. Read it with that in mind.

---

1. MinnsDB

Architecture: Multi-modal memory database. Temporal knowledge graph, vector store, BM25 index, claim store, temporal tables, and structured memory in one Rust binary.

MinnsDB is a memory database for AI agents. MinnsDB stores every fact as a graph edge with a validity window. The valid_from and valid_until timestamps record when a fact became true and when something superseded it. MinnsDB ships an OWL/RDFS ontology layer that drives cascade invalidation. When a user says "I moved to New York," MinnsDB invalidates "lives in London" and cascade-invalidates dependent facts such as "visits the British Museum on weekends." The ontology defines these behaviours in TTL files loaded at startup. No application code is required.

MinnsDB's conversation pipeline injects current graph state into the extraction prompt. Before MinnsDB extracts facts from a new turn, it queries the graph for what it already knows about the entities involved and feeds that state into the LLM. The LLM marks facts as updates rather than producing duplicates the system has to reconcile later.

MinnsDB owns six storage modalities in a single binary:

• Temporal knowledge graph. Edges carry validity windows. Supersession is structural and automatic.
• Vector store. MinnsDB embeds and indexes node vectors, edge vectors over relationship triplets, and claim embeddings with HNSW. No Pinecone, Qdrant, Weaviate, or Chroma required.
• BM25 keyword index. Native lexical search alongside semantic search.
• Claim store. Verified facts carry confidence scores and supersession history.
• Temporal tables. MVCC-versioned relational rows for structured records.
• Structured memory. Ledgers, state machines, preference lists, and tree structures for domain-specific patterns.

MinnsDB Connect ships forty-nine typed source connectors. The connectors cover Slack, Gmail, Jira, Salesforce, GitHub, Google Drive, Notion, and forty-two others. Each source routes through a typed pipeline (events, state changes, transactions, code, documents) and feeds the same temporal graph.

MinnsDB retrieval fuses seven sources via Reciprocal Rank Fusion. The seven sources are BM25 keyword search, node vector similarity, edge vector similarity, verified claims with confidence scores, entity resolution with one-hop graph expansion, DRIFT community search, and episodic recall. Results pass through temporal validity filtering. Only currently-true facts reach the LLM.

MinnsDB runs as a single Rust binary with embedded redb storage. MinnsDB has no external database dependencies. MinnsQL provides a dedicated query language for temporal graph operations. The NLQ endpoint translates natural-language questions into MinnsQL queries with LLM-powered answer synthesis.

Strengths: Six storage modalities in one binary; ontology-driven cascade invalidation; graph-state-aware extraction pipeline; forty-nine source connectors; seven-source retrieval fusion; temporal validity on every fact; LangChain integration via pip install langchain-minnsdb.

Best for: Agents that track state changing over time across many sources. Personalisation, customer success, deal pipelines, project management, and recommendation systems where what is true now differs from what was true last month.

Trade-offs: MinnsDB is a database, not a one-line SDK swap. MinnsDB requires an LLM API key for fact extraction during ingest. For "find the most similar document," a vector store is simpler.

---

2. Zep

Architecture: Temporal knowledge graph (Graphiti) over Neo4j, FalkorDB, or Kuzu.

Zep models memory as a time-aware graph of users, entities, topics, and summaries. Facts carry valid_at and invalid_at dates. Retrieval blends BM25, cosine similarity, and graph traversal, reranked with RRF or MMR. Graphiti, the open-source engine underneath, supports Neo4j, FalkorDB, and Kuzu as storage backends.

Strengths: Low-latency retrieval, production-tested, temporal invalidation, open-source graph engine.

Best for: Chat agents and LLM pipelines that need episodic recall with temporal awareness.

Trade-offs: Zep requires a graph database such as Neo4j as a dependency. Zep does not ship federated source connectors. The graph schema is opinionated for conversational memory.

---

3. Letta (formerly MemGPT)

Architecture: Stateful agent runtime with editable memory blocks, backed by Postgres or SQLite.

Letta makes memory a first-class agent component. Core memory blocks always sit in the prompt. Archival memory lives in a database, and the agent calls tools to retrieve from it. Agents persist across restarts.

Strengths: Transparent and controllable; works well with local models such as vLLM and Ollama; produces genuinely stateful agents.

Best for: Persistent assistants, long-lived workers, and self-hosted local-LLM stacks.

Trade-offs: Letta has no temporal supersession. Memory blocks are edited manually. You own the runtime.

---

4. Cognee

Architecture: LLM-driven memory pipeline producing triplet graphs over Neo4j, with vector and relational stores alongside.

Cognee runs a Remember–Recall–Improve–Forget pipeline. An LLM extracts entities and relations from input data and builds them into a Neo4j graph. The cognify() step constructs the retrieval stack. Auto-routing chooses vector or graph search per query.

Strengths: Flexible custom pipeline composition through task chaining; mature ontology grounding for entity typing; built-in graph visualisation.

Best for: Research workflows, document-heavy agents, and RAG pipelines that need structure beyond similarity search.

Trade-offs: Cognee extraction is stateless. The pipeline does not consult the existing graph while extracting, so updates and contradictions accumulate alongside the facts they should supersede. Cognee requires Neo4j. Cognee has no temporal validity model on extracted facts.

---

5. Anthropic Memory

Architecture: Model-native managed memory built into Claude.

Anthropic Memory lets agents and users write facts and preferences. Claude pulls them into context automatically. Updates and deletes go through the API.

Strengths: Zero infrastructure, deep model integration, privacy-aware by default.

Best for: Teams building on Claude that want memory without operational overhead.

Trade-offs: Anthropic Memory is locked to Claude. Anthropic Memory has no temporal graph and no supersession logic. It is a key-value store with model-level retrieval.

---

6. Mem0

Architecture: Vector store with entity linking and LLM-driven conflict resolution. Backed by Qdrant, Neo4j, and a key-value store.

Mem0 has the largest community in the agent-memory space. Mem0 extracts facts from interactions through an LLM, stores them with vector embeddings, and uses entity linking to boost relevant memories during retrieval. On add(), an LLM-driven conflict-resolution step checks for duplicates and contradictions so the latest claim wins.

Strengths: Polished SDK; broad integrations including LangChain and CrewAI; sensible defaults for personalisation.

Best for: Personalised assistants handling user preferences, profile facts, and conversation summaries.

Trade-offs: Mem0 conflict resolution depends on the LLM correctly identifying contradictions at ingest. Mem0 has no structural model of supersession, no validity windows, and no version history. Updates overwrite, so a question like "what did the user prefer last quarter" cannot be answered.

---

7. LangMem

Architecture: Summarisation-based working memory native to LangGraph.

LangMem is a context compressor more than a memory store. LangMem summarises older turns, re-injects only the relevant summaries, and scopes memory objects to namespaces.

Strengths: Native to LangGraph; few extra moving parts; good for token economics.

Best for: Teams already on LangGraph or LangChain that want memory inside the orchestration layer.

Trade-offs: LangMem is working memory, not long-term memory. Long-range recall is unreliable.

---

8. Supermemory

Architecture: Vector index with temporal metadata.

Supermemory keeps things lightweight: embeddings for semantic recall, time annotations for recency, and ranking that blends both.

Strengths: Simple to deploy; scales like a vector database; time-aware retrieval out of the box.

Best for: Long-running assistants that want recency-weighted semantic recall without graph modelling.

Trade-offs: Supermemory does not model relationships or structured state. For "who did X to whom and when," look elsewhere.

---

9. MemMachine

Architecture: Open-source persistent memory layer with a model-agnostic API.

MemMachine is the community-driven alternative to proprietary memory products. MemMachine captures, normalises, and persists events, then indexes them for semantic or key-based retrieval, all under a permissive licence.

Strengths: Self-hostable, model-agnostic, transparent.

Best for: Teams that need full control over their memory layer or want to avoid vendor lock-in.

Trade-offs: MemMachine sits closer to the metal. Less hand-holding than commercial offerings.

---

10. Memorilabs (Memori)

Architecture: SQL-native memory with relational tables and temporal versioning.

Memori takes a deliberately old-school approach. Structured tables hold facts, entities, and events. Temporal versioning audits how memory evolved. Retrieval is deterministic SQL. Vector embeddings serve as an optional secondary index.

Strengths: Deterministic, auditable, low-cost, multi-tenant friendly.

Best for: Regulated and enterprise environments where governance and explainability matter as much as recall quality.

Trade-offs: Memori is not the right tool for free-flowing conversational memory. The relational schema is your friend and your ceiling.

---

Comparison table

---

How to choose

The market has stopped pretending there is one right answer. The right question is: what does your agent need to remember, and how often does that change?

If your agent's world is mostly static and the job is recall, a vector-based system such as Mem0 or Supermemory is fine. If your agent's world changes, if facts supersede each other, if what is true depends on what changed before it, you need a system that treats memory as a database problem. That is what MinnsDB is built to be: ingestion, extraction, state tracking, invalidation, and retrieval in one stack.