Agent memory: SaaS products and framework modules

Agent memory is the persistence layer that lets an AI agent remember users, facts, and prior sessions beyond a single context window — extraction (“what is worth remembering”), consolidation (“update or contradict what we knew”), and retrieval (“recall the right facts cheaply at inference time”). As of mid-2026 it has hardened from a library choice into a distinct infrastructure category, with dedicated SaaS vendors, hyperscaler managed services, and memory modules embedded in every major agent framework. The wiki has no category or vendor pages for this space yet — this page is a landscape scan filed back from a query (2026-08-24), and a candidate seed for a new taxonomy category.

Dedicated memory SaaS (the shortlist)

Product	Architecture	Deploy	Notes (as of 2026-08)
Mem0	Vector-first, scoped layers (session → user → org)	SaaS + OSS	Category leader by mindshare; $24M Series A (2025-10); free tier then $19–$249/mo; integrates LangChain, CrewAI, LlamaIndex
Zep	Temporal knowledge graph (OSS Graphiti engine, 20k+ GitHub stars)	SaaS + OSS engine	Tracks fact validity over time (edges with validity intervals); Flex plan ~$125/mo (2026-07); best for temporal reasoning
Letta (ex-MemGPT)	Stateful agent runtime; core / recall / archival tiers; agent edits its own memory blocks	SaaS + OSS	UC Berkeley lineage; free self-host, Pro ~$20/mo; more an agent runtime than a bolt-on memory API
Cognee	Graph-native; ECL (Extract–Cognify–Load) pipeline into a knowledge graph	OSS + hosted	Berlin; $7.5M seed (Pebblebed); best-regarded self-hosted graph memory
Supermemory	Universal memory API; dynamic knowledge graphs over user data streams	SaaS	SF; ~$3M raised; founder-led momentum, earlier-stage than the four above
Honcho (Plastic Labs)	“Social memory” — per-user representations, continual learning	SaaS + OSS	Niche: user-modeling / theory-of-mind angle rather than fact recall
Memobase	Profile-based user memory	OSS + SaaS	Smaller entrant; profile/persona memory for consumer-style apps

Open-source & research memory systems (not SaaS)

Frequently name-checked alongside the SaaS shortlist, but these are OSS projects or research artifacts — no hosted product to buy (as of 2026-08):

Project	Who/what	Notes (as of 2026-08)
Hindsight	Vectorize.io OSS (with Virginia Tech, The Washington Post)	Retain / Recall / Reflect model over four memory networks (facts, experiences, entity summaries, beliefs); 91.4% LongMemEval; arXiv paper. Most credible of this group; Vectorize sells the surrounding platform
MemPalace	MIT-licensed OSS, viral 2026-04 (19.5k GitHub stars in week one)	“Memory palace” hierarchical index (wings/rooms/drawers), verbatim storage, fully local, MCP integration; claims 96.6% LongMemEval — independent commentary flags the benchmark method as inflated (retrieves near-whole dataset)
Memora	Microsoft Research (ICML 2026 paper + OSS code)	“Harmonic” representation decoupling storage from retrieval; up to 98% context-token reduction claimed. Research framework, not a product; a separate small OSS project shares the name (agentic-box/memora)
OpenBrain	Small local-first OSS	Personal memory over Postgres/pgvector + Ollama, exposed as an MCP server; provider-agnostic, no cloud dependency. Early/hobbyist-scale

Hyperscaler managed memory

Product	Status	Notes
AWS Bedrock AgentCore Memory	GA since 2025-10	Short-term session + long-term retention; managed or self-managed extraction strategies; episodic memory at GA; works with LangGraph, CrewAI, Strands. The most operationally complete managed offering
Google Vertex AI Memory Bank	GA (with Sessions)	HIPAA support; $0.25 per 1,000 session/memory events; part of Agent Engine
cloudflare Agent Memory	Private beta (2026-04)	Edge-native; pricing unpublished

Framework/vendor-embedded memory modules

Not products you buy separately, but the module you get “for free” with a stack — and the main reason many teams never buy a dedicated memory SaaS:

LangMem (LangChain/LangGraph) — hot-path memory tools plus background managers that extract, consolidate, and search long-term memory. Default choice if you’re already on LangGraph.
CrewAI memory — unified memory API with LLM-assisted scope/category/ importance inference.
LlamaIndex memory — composable short/long-term memory blocks over its retrieval stack.
OpenAI memory — user-level memory in ChatGPT and the Assistants/Responses stack; closed, tied to the platform.
Anthropic Claude memory tool — file-based memory directory the agent reads/ writes via the Agent SDK; app-controlled persistence.
Redis Agent Memory Server — low-latency working + long-term memory backend; best when Redis is already in the estate.
pinecone and other vector DBs — the storage substrate many of the above sit on, not a memory manager by themselves.

How to read the space

The common stack is two layers: a memory manager (Mem0/Zep/Cognee/LangMem) for extraction + consolidation, over a storage layer (Pinecone, Redis, pgvector).
Vector vs. graph is the main architectural split: vector-first (Mem0) optimizes cheap similarity recall; graph/temporal (Zep, Cognee, Supermemory) optimizes entity- and time-aware reasoning (“what did we know, when”).
Runtime vs. sidecar: Letta wants to be the agent runtime; Mem0/Zep bolt onto whatever runtime you have. Different procurement decisions.
For a fund CTO, the governance angle: memory stores are a new repository of client/user data outside the document estate — retention, entitlement, and DLP controls rarely reach them yet. Day-2 concern; adjacent to entitlement-aware RAG.

Open questions

No agent-memory category exists in taxonomy.yaml; no vendor pages for Mem0, Zep, Letta, Cognee, Supermemory. Decide whether this earns a category (likely emerging tier) or stays a comparison page.
Benchmark claims (Mem0’s 92.5% LoCoMo, 94.4% LongMemEval) are vendor-reported — unverified against independent runs.
Pricing is volatile; all figures dated 2026-07/08 from secondary sources, not cached primary pages.

Sources

Accessed 2026-08-24 via web search (secondary/roundup-heavy; not yet cached to raw/sources/ — do that if any claim gets promoted to a vendor page):

AI Governance & Security Vendor Wiki

Explorer