Agent memory - verified
We read the docs, pricing pages, READMEs and license files of 15 agent memory products on October 18, 2026: 4 hosted memory APIs, 5 open source frameworks, 4 cloud agent platform features and 2 model-provider tools. Every value carries a verbatim quote and its source. Prices stay in each product's own unit, so they are grouped by unit and never ranked across units.
Want the long read? See Mem0 vs Zep vs Letta vs Cloudflare or agent memory architectures explained. Storing embeddings yourself? Compare vector databases.
Editions no longer supported, servers no longer maintained, moved repositories, retired APIs, replaced graph stores, a deprecated backend and new billing, each said only as far as the product's own page says it, with the quote and the link. Where the page gives no date, neither do we.
Zep's README says Zep Community Edition is no longer supported. Zep's pricing page describes deploying Zep inside your own VPC. No date given.
“Zep Community Edition is no longer supported.”
“Deploy Zep inside your VPC. Your network, your perimeter, your compliance boundary.”
The README of Zep's earlier Go MCP server says it is no longer maintained. Zep's docs now describe a Context MCP Server. No date given.
“This MCP server is no longer maintained.”
“The Context MCP Server lets an end user connect an MCP client (Claude, ChatGPT, Claude Code, Codex, Cursor, and others) to their own agent memory in Zep.”
The letta-ai/letta README says the current source code lives in letta-ai/letta-code. No date given.
“The current source code lives in letta-ai/letta-code, which includes the agent harness, interactive terminal UI, App Server, channels, and the runtime used by the desktop and web apps.”
The same README says the archive branch contains the retired Letta V1 API server. No date given.
“The archive branch contains the retired Letta V1 API server.”
Letta Code's README says AgentFile (.af) export and import are deprecated and have been removed from Letta Code. No date given.
“AgentFile (.af) export and import are deprecated and have been removed from Letta Code.”
Mem0's graph memory docs say the external graph store integration has been replaced by native, built-in Graph Memory on the platform, with no Neo4j, Memgraph or other graph store to deploy. No date given.
“That integration has been replaced by native, built-in Graph Memory:”
“Graph Memory is built in. There is no Neo4j, Memgraph, or other graph store to deploy, no connection strings to manage, and nothing to enable. It runs natively inside the platform and is always on.”
Mem0's Platform vs open source page says the open-source graph store integration was dropped when the v3 pipeline landed. The open-source SDK still extracts entities to boost ranking, but has no queryable graph. No date given.
“that integration was dropped when the v3 pipeline landed”
“OSS still extracts entities and uses them to boost ranking, but there is no queryable graph and no relations field”
Graphiti's README lists its Kuzu 0.11.2 backend as deprecated. No date given.
“Kuzu 0.11.2 (deprecated, see below)”
The agent-memory-server README says Redis Agent Memory is part of Redis Iris, Redis's real-time context engine for agents. The original Redis Agent Memory Server now sits in V0/, kept as an open research artifact rather than the supported production distribution. No date given.
“It is part of Redis Iris, Redis's real-time context engine for agents.”
“V0/ is the original Redis Agent Memory Server. It informed the Iris product.”
“We keep it here as an open research artifact, not as the supported production distribution.”
Cognee's README says using Postgres as a graph store is released as a demo feature, and that the production-ready version is a licensed product. No date given.
“Using Postgres as a graph store is currently released as a demo feature. The production-ready version is available as a licensed product.”
Google's Agent Platform pricing page says Memory Bank billing on its current pricing structure will commence on September 1st, 2026.
“Memory Bank billing on this pricing structure will commence on September 1st, 2026.”
AWS's AgentCore pricing page shows short-term memory ingestion at $0.25 per 1,000 new events through October 5, 2026, and $1.00 per GB ingested as of October 6, 2026.
“$0.25 per 1,000 new events (Through October 5, 2026) $1.00 per GB ingested (As of October 6, 2026)”
“$1.00 per GB ingested (As of October 6, 2026)”
Motorhead's README says support is no longer maintained for the project. Not included in this directory. No date given.
“Support is no longer maintained for this project.”
Each card shows how the product charges in its own words, then its memory types, knowledge graph, automatic extraction, MCP server, self-hosting and license. Every value links the page it came from.
Managed services you send conversations to and query for memories, priced per plan, credit, token or operation.
How it charges
Free - Hobby planinferred
“Price Free $19/month $249/month Custom”
How it charges
Start free with 10,000 credits a month - credits per month (Free plan)inferred
“Prototyping? Start free with 10,000 credits a month.”
How it charges
0 USD with 5 USD included credits - Free planinferred
“Free · 0 USD · 5 USD · Plugins, API evaluation, prototypes”
How it charges
Not statedNot stated in the repository README; Redis publishes pricing with Redis Cloud, which is outside the pages this record cites.
Libraries and servers you run yourself, with a license linked. Some add a hosted tier beside the code.
How it charges
No usage-based billing (self-hosted) - n/a
“Cost optimization: local LLMs (Ollama), self-hosted vector DBs, no usage-based billing.”
How it charges
Not statedOpen-source framework; the README states no price.
How it charges
Not statedOpen-source library; the README states no price.
Memory built into a cloud provider's agent platform, billed in that platform's units and tied to its runtime.
How it charges
$0.30 per GiB-month - Agent Storage, GiB-month
“Storage: Total data stored (including revisions), billed as Agent Storage ($0.30/GiB-month).”
Every value, quote and source for Agent Platform Memory Bank
How it charges
$1.00 per GB ingested (as of October 6, 2026) - short-term memory ingestion, per GB
“$1.00 per GB ingested (As of October 6, 2026)”
Every value, quote and source for Amazon Bedrock AgentCore Memory
How it charges
Billed for underlying chat and embedding model usage - model usage
“You're billed for usage of the underlying chat and embedding models you configure.”
Every value, quote and source for Memory in Microsoft Foundry Agent Service
How it charges
Not billed during private beta - n/a
“We are not currently billing for Agent Memory during private beta.”
Memory that ships with a model provider's API or agent SDK, where your own application stores the data.
How it charges
Client-side tools are priced the same as any other Claude API request - model tokens (the memory tool is a client tool)
“Client-side tools are priced the same as any other Claude API request”
How it charges
Not statedNot stated on the SDK sessions page (the SDK is open source; model and Conversations API usage are billed by OpenAI elsewhere).
Every price line, in the product's own words, grouped by what it charges for. Rows are sorted by amount only inside one unit; a plan, a credit, an active agent, a memory token, an operation and a GB are different things, so nothing is converted or ranked across units.
Flat plan prices, per month where the product says so, with free and custom plans beside them.
| Product | Price, in the product's words | Quote and source |
|---|---|---|
| Unit: Hobby plan | ||
| Mem0 Platform | Freeinferred | “Price Free $19/month $249/month Custom” |
| Unit: per month (Starter plan) | ||
| Mem0 Platform | $19inferred | “Price Free $19/month $249/month Custom” |
| Unit: per month (Pro plan) | ||
| Mem0 Platform | $249inferred | “Price Free $19/month $249/month Custom” |
| Unit: Enterprise plan | ||
| Mem0 Platform | Custominferred | “Price Free $19/month $249/month Custom” |
| Zep | Custom credits with negotiated ratesinferred | “Custom credits with negotiated rates” |
| Supermemory | Custom | “Enterprise · Custom · Contract · SSO, custom metering, FDE, air-gap” |
| Unit: per month, billed through the Vercel Marketplace integration | ||
| Mem0 Platform | Hobby (free) or Starter ($20/month) | “Select Hobby (free) or Starter ($20/month).” |
| Unit: Free plan | ||
| Supermemory | 0 USD with 5 USD included creditsinferred | “Free · 0 USD · 5 USD · Plugins, API evaluation, prototypes” |
| Letta | $0/monthinferred | “$0 /month” |
| Cognee | $0/monthinferred | “$0/month” |
| Unit: Pro plan | ||
| Supermemory | 19 USD/mo with 20 USD included creditsinferred | “Pro · 19 USD/mo · 20 USD · Developers, core connectors” |
| Unit: Max plan | ||
| Supermemory | 100 USD/mo with 130 USD included creditsinferred | “Max · 100 USD/mo · 130 USD · More usage headroom and Gmail” |
| Unit: Scale plan | ||
| Supermemory | 399 USD/mo with 600 USD included creditsinferred | “Scale · 399 USD/mo · 600 USD · Production, full connectors, teams” |
| Unit: per month (Pro, personal) | ||
| Letta | $20/monthinferred | “$20 /month” |
| Unit: Enterprise | ||
| Cognee | BYOC engagementinferred | “Delivered as a BYOC engagement: your own cloud, dedicated support, and SLAs for memory at scale.” |
Zep bills in credits: a monthly allowance per plan, then credits per Episode by size.
| Product | Price, in the product's words | Quote and source |
|---|---|---|
| Unit: credits per month (Free plan) | ||
| Zep | Start free with 10,000 credits a monthinferred | “Prototyping? Start free with 10,000 credits a month.” |
| Unit: per month (Flex plan), credits | ||
| Zep | $125/month including 50,000 credits; then $25 per 10,000 creditsinferred | “50,000 credits included. Then $25 / 10,000 credits.” |
| Unit: per month (Flex Plus plan), credits | ||
| Zep | $375/month including 200,000 credits; then $75 per 40,000 creditsinferred | “200,000 credits included. Then $75 / 40,000 credits.” |
| Unit: credits per Episode | ||
| Zep | 1 credit per Episode up to 350 bytes; +1 credit per additional 350 bytes | “Episodes up to 350 bytes use 1 credit; each additional 350 bytes (or part thereof) uses another credit.” |
Usage priced per active agent, per seat, per extra workspace or per second of tool execution. Different units; never converted.
| Product | Price, in the product's words | Quote and source |
|---|---|---|
| Unit: per active agent per month (listed under the API Plan) | ||
| Letta | $0.10 per active agent per monthinferred | “$0.10 / active agent / mo” |
| Unit: per second (listed under the API Plan) | ||
| Letta | $0.00015 per second of tool executioninferred | “$0.00015 / sec tool execution” |
| Unit: per seat per month (Teams Pro) | ||
| Letta | $20 per seat per monthinferred | “$20 /seat/month” |
| Unit: per workspace (Standard) | ||
| Cognee | $5 per additional workspaceinferred | “+ $5 per additional workspace” |
Memory tokens priced by the product, or the model tokens a memory feature consumes and the provider bills elsewhere.
| Product | Price, in the product's words | Quote and source |
|---|---|---|
| Unit: memory tokens (text) | ||
| Supermemory | 0.005 USD per 1K tokens | “sm_tokens_text · 0.000005 · 0.005 USD / 1K tokens” |
| Unit: memory tokens (rich) | ||
| Supermemory | 0.010 USD per 1K tokens | “sm_tokens_rich · 0.00001 · 0.010 USD / 1K tokens” |
| Unit: tokens | ||
| Letta | LLM usage at underlying model token cost | “LLM usage is charged based on the underlying token costs of the model used.” |
| Unit: per 1M tokens (Standard) | ||
| Cognee | $1.00 per 1M tokensinferred | “$1.00/ 1M tokens” |
| Unit: model SKUs | ||
| Agent Platform Memory Bank | Model tokens billed separately | “Tokens: model tokens used for memory generation and embeddings are billed separately under their respective model SKUs.” |
| Unit: model usage | ||
| Memory in Microsoft Foundry Agent Service | Billed for underlying chat and embedding model usage | “You're billed for usage of the underlying chat and embedding models you configure.” |
| Unit: model tokens (the memory tool is a client tool) | ||
| Claude memory tool | Client-side tools are priced the same as any other Claude API request | “Client-side tools are priced the same as any other Claude API request” |
Usage priced per thousand or million read, write, search or retrieval operations, each in its own unit.
| Product | Price, in the product's words | Quote and source |
|---|---|---|
| Unit: search queries | ||
| Supermemory | 0.005 USD per 1K queries | “sm_search_queries · 0.000005 · 0.005 USD / 1K queries” |
| Unit: operations | ||
| Supermemory | 0.10 USD per 1K operations | “sm_operations · 0.0001 · 0.10 USD / 1K operations” |
| Unit: read operations | ||
| Agent Platform Memory Bank | 1 Agent Compute vCPU-h ($0.085) per 3 million read operations | “Read API requests: 1 Agent Compute vCPU-h ($0.085) for every 3 million read operations.” |
| Unit: write operations | ||
| Agent Platform Memory Bank | 1 Agent Compute vCPU-h ($0.085) per 1 million write operations | “Write API requests: 1 Agent Compute vCPU-h ($0.085) for every 1 million write operations.” |
| Unit: long-term memory retrieval | ||
| Amazon Bedrock AgentCore Memory | $0.50 per 1,000 retrievals | “Long-Term Memory Retrieval · $0.50 per 1,000 retrievals” |
Usage priced per GB ingested or retrieved, or per GB-month or GiB-month stored. A GB is not a GiB; nothing is converted.
| Product | Price, in the product's words | Quote and source |
|---|---|---|
| Unit: Agent Storage, GiB-month | ||
| Agent Platform Memory Bank | $0.30 per GiB-month | “Storage: Total data stored (including revisions), billed as Agent Storage ($0.30/GiB-month).” |
| Unit: short-term memory ingestion, per GB | ||
| Amazon Bedrock AgentCore Memory | $1.00 per GB ingested (as of October 6, 2026) | “$1.00 per GB ingested (As of October 6, 2026)” |
| Unit: short-term memory retrieval, per GB | ||
| Amazon Bedrock AgentCore Memory | $0.20 per GB retrieved | “$0.20 per GB retrieved (As of October 6, 2026)” |
| Unit: short-term memory storage, per GB-month | ||
| Amazon Bedrock AgentCore Memory | $0.10 per GB-month | “$0.10 per GB-month (As of October 6, 2026)” |
Long-term memory storage priced per 1,000 records per month.
| Product | Price, in the product's words | Quote and source |
|---|---|---|
| Unit: long-term memory storage (built-in strategies) | ||
| Amazon Bedrock AgentCore Memory | $0.75 per 1,000 records/month | “Long-Term Memory Storage (built-in) · $0.75 per 1,000 records/month” |
| Unit: long-term memory storage (override/self-managed strategies) | ||
| Amazon Bedrock AgentCore Memory | $0.25 per 1,000 records/month | “Long-Term Memory Storage (override/self-managed) · $0.25 per 1,000 records/month*” |
How billing works where there is no rate: free tiers, billing granularity, usage pricing on request and preview or beta terms.
| Product | Price, in the product's words | Quote and source |
|---|---|---|
| Unit: on request | ||
| Mem0 Platform | Usage-based pricing available | “We also support usage-based pricing for teams whose traffic doesn’t map cleanly to a fixed tier.” |
| Unit: n/a | ||
| Mem0 Open Source | No usage-based billing (self-hosted) | “Cost optimization: local LLMs (Ollama), self-hosted vector DBs, no usage-based billing.” |
| Memory in Microsoft Foundry Agent Service | Pricing may change during preview | “Pricing and billing for memory and the Memory Store API can change during preview.” |
| Cloudflare Agent Memory | Not billed during private beta | “We are not currently billing for Agent Memory during private beta.” |
| Cloudflare Agent Memory | At least 30 days notice before charging | “We will provide at least 30 days notice before we make any changes or start charging for Agent Memory usage.” |
| Unit: free tier | ||
| Agent Platform Memory Bank | Monthly free usage tier per account for each resourceplatform-wide | “A free usage tier applies per month, per account for each resource.” |
| Unit: billing granularity | ||
| Amazon Bedrock AgentCore Memory | Each event billed at a minimum of 12 KB and a maximum of 64 KB | “Each event is billed at a minimum of 12 KB and a maximum of 64 KB, regardless of its actual size.” |
| Unit: AWS Free Tier credits | ||
| Amazon Bedrock AgentCore Memory | Up to $200 in Free Tier credits for new AWS customersplatform-wide | “New AWS customers receive up to $200 in Free Tier credits.” |
Two of Mem0's pages say different things. Both quotes are shown; we do not pick one.
Mem0's pricing page lists $19/month in the Starter column; Mem0's Vercel Marketplace integration doc says Starter is $20/month. Both are shown as stated and not reconciled.
Vercel integration doc
Hobby (free) or Starter ($20/month)
“Select Hobby (free) or Starter ($20/month).”
Mem0's graph memory docs say graph memory runs on all plans, automatically; the pricing page lists Graph memory (entity linking) in the Pro and Enterprise plan cards. Both are shown as stated.
Graph memory docs
Graph memory on all plans (docs)
“Graph memory: entity extraction, linking, and the retrieval boost · All plans, automatic”
Pricing page
Pricing page lists Graph memory (entity linking) in the Pro and Enterprise plan cardsinferred
“Graph memory (entity linking)”
Every product against every field, each cell linking its source. Not stated is not the same as no. This is not a ranking.
Memory types and scopes, where memories are stored, knowledge graphs, how they are retrieved and whether extraction is automatic.
Official MCP servers, SDKs, self-hosting and license, retention and deletion controls, and compliance as stated.
Each product's launch stage exactly as its page labels it. One status is inferred, and says from what. No label is shown as Not stated.
| Product | Launch stage | Quote and source |
|---|---|---|
| Mem0 Platform | Not statedNo launch-stage label (GA/beta) on the pricing page or the platform overview page. | |
| Zep | Not statedNo launch-stage label on the pricing page or the MCP server page. | |
| Supermemory | Not statedNo launch-stage label on the docs pages reviewed. | |
| Redis Agent Memory | Not statedNo launch-stage label in the README. | |
| Mem0 Open Source | Not statedNo launch-stage label in the README or the platform-vs-OSS page. | |
| Graphiti | MCP server described as experimental (MCP server component only) | “This is an experimental Model Context Protocol (MCP) server implementation for Graphiti.” |
| Letta | Actively developed | “Letta (f.k.a. MemGPT) is actively developed.” |
| LangMem | Not statedNo launch-stage label in the README. | |
| Cognee | Not statedNo launch-stage label on the pricing page or README. | |
| Agent Platform Memory Bank | Not statedNo launch-stage label on the Memory Bank overview page. | |
| Amazon Bedrock AgentCore Memory | Not statedNo launch-stage label on the memory overview page. | |
| Memory in Microsoft Foundry Agent Service | Public preview | “Memory is currently in public preview.” |
| Cloudflare Agent Memory | Private beta | “Agent Memory is in private beta.” |
| Claude memory tool | Generally availableinferredInferred from "doesn't require a beta header"; the page gives no GA label, and the SDK helpers still live in a beta namespace. | “The helper and tool-runner surfaces live in your SDK's beta namespace even though the memory tool itself doesn't require a beta header.” |
| OpenAI Agents SDK memory | Not statedNo launch-stage label on the sessions page. |
Benchmark and latency figures as each vendor states them about itself, under its own harness. We did not measure them, and they do not compare across vendors.
Vendor-stated: 92.5 on LoCoMo, +21 points over the previous algorithm
“92.5 on LoCoMo -- +21 points over the previous algorithm”
Vendor-stated: 94.4 on LongMemEval
“94.4 on LongMemEval -- +27 points, with 98.2 on assistant memory recall”
Vendor note: scores reflect the managed platform
“Scores reflect Mem0's managed platform, which includes proprietary optimizations not available in the open-source SDK”
Vendor-stated: sub-200ms retrieval at scale
“Pre-configured, production-ready retrieval with sub-200ms performance at scale”
Vendor-stated: leads LongMemEval and LoCoMo
“Supermemory leads the LongMemEval and LoCoMo benchmarks and independent ones such as SWEContext.”
Vendor note: OSS users should expect directionally similar but not identical benchmark numbers
“open-source users should expect directionally similar gains but not identical numbers”
Vendor-stated: 0.79 at 100K tokens
“100K tokens · 0.79 · Fixed hybrid retrieval; four evaluation rounds over 20 questions from one held-out conversation.”
Products whose own pages state each capability, in directory order. These are filters, not picks.
The product's own pages say it builds a graph of entities or facts.
An MCP server from the vendor's own docs or repository; check whether it is hosted or experimental.
Code you can run yourself with a license quoted. Supermemory is left out: its self-hosted server binary is not open source.
Redis Agent Memory, Mem0 Open Source, Graphiti, Letta, LangMem, Cognee, OpenAI Agents SDK memory
Memories are extracted from conversations or runs without you writing each one.
Mem0 Platform, Zep, Supermemory, Redis Agent Memory, Mem0 Open Source, Graphiti, Letta, LangMem, Cognee, Agent Platform Memory Bank, Amazon Bedrock AgentCore Memory, Memory in Microsoft Foundry Agent Service, Cloudflare Agent Memory, Claude memory tool, OpenAI Agents SDK memory
Delete, expire or TTL controls described on the product's own pages.
Mem0 Platform, Zep, Supermemory, Redis Agent Memory, Mem0 Open Source, Graphiti, Cognee, Agent Platform Memory Bank, Amazon Bedrock AgentCore Memory, Memory in Microsoft Foundry Agent Service, Cloudflare Agent Memory, Claude memory tool, OpenAI Agents SDK memory
SOC 2, HIPAA, GDPR or DPA statements on the vendor's own pages.
Values whose reach is wider or narrower than they look: free tiers that cover a whole cloud platform, and a self-hosted server that is not open source even though its repository carries an MIT license.
Monthly free usage tier per account for each resource
This free-tier line is on the Agent Platform pricing page and applies per month, per account, for each resource. It is not specific to Memory Bank.
“A free usage tier applies per month, per account for each resource.”
Up to $200 in Free Tier credits for new AWS customers
AWS Free Tier credits for new AWS customers, stated on the AgentCore pricing page. Not specific to AgentCore Memory.
“New AWS customers receive up to $200 in Free Tier credits.”
Yes, self-hosted binary (server built from a non-public codebase)
The self-hosted server binary is built from a separate, non-public codebase, so the self-host option is not open source. The repository LICENSE is MIT and the SDK packages declare Apache-2.0.
“The SDKs and other components in the public repository are open source; the downloadable self-hosted server binary is built from a separate, non-public codebase.”
MIT (public repository)
Covers the public repository (SDKs and other components), not the downloadable self-hosted server binary.
“MIT License”
Values whose plan name, table column or language comes from the page layout rather than the quote, each tagged inferred wherever it renders.
Products considered and left out, with the reason and, where one exists, the quote recorded at the time.
Motorhead (getmetal/motorhead)
Deprecated and unmaintained per its own README.
“Support is no longer maintained for this project.”
MongoDB agent memory
No dedicated MongoDB agent memory product page was found among MongoDB's own primary sources; MongoDB appears only as a storage backend in other products' docs (for example OpenAI's MongoDBSession).
OpenAI API server-side memory
No standalone memory API for developers was found in OpenAI's own SDK docs; the Agents SDK sessions and sandbox-agent memory are recorded instead. ChatGPT consumer memory is out of scope.
Our longer reads on memory architectures, patterns, benchmarks and the products in this directory. Point-in-time posts; where a post and this directory differ, the directory is the newer read.
Method
Read from each product's docs, pricing pages, READMEs and license files on , with every quote string-matched against the page text that day. Prices, plans and support status change; check the linked source before you commit.
In the vendors' own words: Zep: "Zep Community Edition is no longer supported." Zep: "This MCP server is no longer maintained." Letta: "The current source code lives in letta-ai/letta-code, which includes the agent harness, interactive terminal UI, App Server, channels, and the runtime used by the desktop and web apps." Letta: "The archive branch contains the retired Letta V1 API server." Letta: "AgentFile (.af) export and import are deprecated and have been removed from Letta Code." Mem0 Platform: "That integration has been replaced by native, built-in Graph Memory:" Mem0 Open Source: "that integration was dropped when the v3 pipeline landed" Graphiti: "Kuzu 0.11.2 (deprecated, see below)" Redis Agent Memory: "It is part of Redis Iris, Redis's real-time context engine for agents." Cognee: "Using Postgres as a graph store is currently released as a demo feature. The production-ready version is available as a licensed product." Agent Platform Memory Bank: "Memory Bank billing on this pricing structure will commence on September 1st, 2026." Amazon Bedrock AgentCore Memory: "$0.25 per 1,000 new events (Through October 5, 2026) $1.00 per GB ingested (As of October 6, 2026)" Only Google (September 1st, 2026) and AWS (October 6, 2026) give a date on the pages we read.
No. Zep's README says: "Zep Community Edition is no longer supported." Self-hosting is described on Zep's pricing page as deploying Zep inside your VPC, and Graphiti, Zep's open-source graph framework, is self-hosted only under the Apache License 2.0.
5 products state one on its own pages: Mem0 Platform, Zep, Supermemory, Graphiti and Cognee. Mem0's open-source SDK extracts entities to boost ranking but, per its docs, has no queryable graph. Not stated elsewhere, which is not the same as unsupported.
5 products describe an MCP server in its own docs or repository: Mem0 Platform (official MCP server hosted by Mem0), Zep (official hosted Context MCP Server), Supermemory (Supermemory MCP server), Graphiti (experimental MCP server in the Graphiti repository) and Cognee (cognee-mcp server in the Cognee repository).
Partly. Its repository LICENSE is MIT and its SDK packages declare Apache-2.0, but its self-hosting docs say: "The SDKs and other components in the public repository are open source; the downloadable self-hosted server binary is built from a separate, non-public codebase." So the self-hosted server is not open source.
In different units, so the prices do not rank: plan prices (Mem0 Platform, Zep, Supermemory, Letta and Cognee); credits (Zep); per agent, seat, workspace or second (Letta and Cognee); tokens and model usage (Supermemory, Letta, Cognee, Agent Platform Memory Bank, Memory in Microsoft Foundry Agent Service and Claude memory tool); requests, queries and operations (Supermemory, Agent Platform Memory Bank and Amazon Bedrock AgentCore Memory); data volume (Agent Platform Memory Bank and Amazon Bedrock AgentCore Memory); stored records (Amazon Bedrock AgentCore Memory). Every price line is listed in the product's own unit with its quote.
No. Every value is a first-party statement copied from the product's own pages, with the quote and the link. Nothing here measures recall accuracy or latency; benchmark scores and latency figures appear only under a Vendor-stated label, and prices in different units are never converted or compared.
Not stated means the pages we read did not say; it is not evidence the feature is missing. Vendor-stated marks a benchmark or latency claim the vendor makes about itself, unverified. Inferred marks our reading: a plan name taken from a pricing table's layout, a column assigned by table order, or a status the page implies without labelling it.
Last verified from each product's own docs, pricing pages, READMEs and license files; every quote was string-matched against the page that day. This is not a benchmark. Prices are shown in the product's own unit and change without notice, so check the linked source before you commit.
Snapshot of October 18, 2026: 15 products, 263 quotes, 13 changes, 3 considered and not included.
Newsletter
One email per week. New model releases, agent tooling changes and what they mean for what you build.
The stores many memory layers sit on: pricing in each vendor's unit, hybrid search, built-in embedding and licenses.
LangGraph, OpenAI Agents SDK, Google ADK and more: the agents these memory layers plug into.
Trace and evaluate what your agent remembers and retrieves, tool by tool.
Where agents run the code they write: isolation, session limits and persistence.