Best of · Agent runtime
Best memory layers for AI agents
All 10 ranked memory layers on the Anchor benchmark, with a pick for each need and where each one falls short. Scores come from public evidence, re-checked as vendors change.
- 10 ranked
- 1 agent-ready
- 1 accept x402
- 7 hosted endpoints
- Updated 8 October 2026
Top three
Picks by need
Worked out from the scores, prices and facts, so they change when the research does.
Highest score overall
Amazon Bedrock AgentCore Memory A
A, 79.4/100 on the benchmark.
Also Zep, B, 69.3/100.
Maintenance & community
85/100 on maintenance & community, against 83 for the overall leader.
Self-hosting under an open licence
Honcho B
self-hosted, AGPL-3 licence.
Also Supermemory API + MCP, self-hosted, MIT licence.
The shortlist
| # | Tool | Grade | Best for | Price | Where |
|---|---|---|---|---|---|
| 1 | Amazon Bedrock AgentCore Memory Amazon Web Services |
A 79.4 | Teams already on AWS that want per-user memory under IAM, KMS and Regional controls, with extraction run for them. | $1 / GB | hosted and local |
| 2 | Zep Zep |
B 69.3 | Agents that must track how facts about a person or account change over time, inside a team that needs access policies and audit logs. | $125 / mo | hosted |
| 3 | Honcho Plastic Labs |
B 64.1 | Products that model the people in a conversation and want to ask questions about them, and for agents that must pay their own way. | $0.001 / call | hosted |
| 4 | Supermemory API + MCP Supermemory |
B 63.6 | Apps that need user memory and document search over the same data, and for multi-tenant products that want a key per user. | $19 / mo | hosted |
| 5 | Mem0 Platform + MCP Mem0 |
C 56.1 | Chat products that want per-user facts back with one search call and little setup. | $19 / mo | hosted |
| 6 | Cognee Cognee |
C 54.6 | Agents whose memory has to include documents, wikis and chat tools as well as conversation, and for teams happy to self-host. | $5 / mo | hosted and local |
| 7 | Memory (MCP reference server) MCP project (reference servers) |
C 54.2 | A single local agent that wants a small, inspectable store of facts about people and projects. | Free · OSS | local |
| 8 | Graphiti Zep |
D 53.3 | Teams that want Zep's temporal graph model on their own infrastructure and can run a graph database. | Free · OSS | library |
| 9 | Hindsight Vectorize |
D 50.2 | Agents that should form opinions and summaries from what they stored, such as long-running assistants or copilots that reflect on past sessions. | $0.05 / call | hosted |
| 10 | LangMem LangChain |
D 47.9 | Teams already on LangGraph that want memory tools and background extraction over a store they run. | Free · OSS | local |
How to choose
- Recall of early factsCheck whether facts from early sessions come back after many later ones, since a memory layer that forgets the first sessions can hand the agent stale context.
- Updates to changed preferencesCheck that a changed preference replaces the old value rather than sitting beside it, because an agent that recalls both may act on the outdated one.
- Deletion that sticksCheck that a deleted fact stops appearing in recall and whether copies remain in logs or derived indexes, since a fact that returns after deletion is a privacy failure.
- Where memory is stored and encryptedCheck where the memory store is hosted, how it is encrypted and whether the vendor trains on what you send, because a user's history is personal data.
How the benchmark tests this category. One user's history fed in over several sessions, then questions that need facts from early on, a changed preference and a deleted fact. We check what comes back, how fast, and whether the deletion sticks.
Each one in detail
Amazon Bedrock AgentCore Memory
A 79.4/100Managed memory service for AI agents on AWS. It stores conversation events as short-term memory and extracts facts, preferences, summaries and episodes into searchable long-term records, through the AWS API, SDKs and an MCP server.
Verdict Access is IAM-controlled down to one namespace, with published per-second quotas for every operation and a clientToken on event writes. Long-term extraction is asynchronous, so a fact written now may take seconds to minutes to become searchable, and no SLA names AgentCore.
Choose it for Teams already on AWS that want per-user memory under IAM, KMS and Regional controls, with extraction run for them.
Strengths
- IAM permissions per operation, resource-based policies and namespace condition keys, with OAuth sign-in and deny-by-default Cedar policies available through AgentCore Gateway
- Per-second quotas published for every Memory operation, such as 200
CreateEventand 30RetrieveMemoryRecordsrequests a second per account and Region CreateEventtakes aclientToken, so a retried write is ignored instead of stored twice
Weaknesses
- Long-term extraction is asynchronous. The docs say records appear within seconds to minutes after a write
- No SLA names AgentCore. The Bedrock SLA of 4 October 2023 covers the Bedrock APIs for models
- Built-in strategies use cross-Region inference, so event text can be processed in another Region of the same geography
Price $1 / GBAuth OAuth or keyx402 nohosted and local
Zep
B 69.3/100Hosted context and memory service built on a temporal graph of facts, relationships and source episodes.
Verdict Facts that new data invalidates keep the time they stopped being true. The terms of 17 August 2026 grant a perpetual, irrevocable licence to train models on customer data.
Choose it for Agents that must track how facts about a person or account change over time, inside a team that needs access policies and audit logs.
Strengths
- Facts that new data invalidates keep the time they stopped being true
- API keys can carry ABAC policies, and audit and API logs record what each key did
- Published rate limits (600 a minute on Flex, 1,000 on Flex Plus) with 429, Retry-After and X-RateLimit headers
Weaknesses
- The terms of 17 August 2026 grant a perpetual, irrevocable licence to train models on customer data
- First paid plan is $125 a month, and credits grow with episode size
- The MCP server needs OAuth through a company identity provider, so a headless agent can't use it
Price $125 / moAuth OAuth or keyx402 nohosted
Full assessment · Against #1, Amazon Bedrock AgentCore Memory
Honcho
B 64.1/100Memory API that models each participant (a peer) in a conversation.
Verdict Pay per call over x402 or MPP at agentcash.honcho.dev, with free read endpoints. No published rate limits, 429 guidance or SLA.
Choose it for Products that model the people in a conversation and want to ask questions about them, and for agents that must pay their own way.
Strengths
- Pay per call over x402 or MPP at agentcash.honcho.dev, with free read endpoints
- Keys can be minted per workspace, peer or session, with an expiry, through the API
- Context retrieval listed as unlimited, with reasoning billed only when you call chat
Weaknesses
- No published rate limits, 429 guidance or SLA
- Changelog entries carry no dates and the repo has no GitHub releases
- The MCP tool list is sent on connect, not documented
Price $0.001 / callAuth OAuth or keyx402 yeshosted
Full assessment · Against #1, Amazon Bedrock AgentCore Memory
Supermemory API + MCP
B 63.6/100Memory and context API that ingests text, URLs, PDFs, images and video, extracts memories into a graph per container tag (usually one per user) and returns them through search or a user profile endpoint.
Verdict 8-tool hosted MCP with OAuth and per-space read or write permission. No legal entity named in the terms or privacy policy.
Choose it for Apps that need user memory and document search over the same data, and for multi-tenant products that want a key per user.
Strengths
- 8-tool hosted MCP with OAuth and per-space read or write permission
- Scoped keys limited to container tags, with expiry from 1 to 365 days
- Status page shows 100 per cent for the API and Console from July to October 2026
Weaknesses
- No legal entity named in the terms or privacy policy
- No published rate limits or 429 guidance
- Ingests web pages and PDFs with no prompt-injection guidance found
Price $19 / moAuth OAuth or keyx402 nohosted
Full assessment · Against #1, Amazon Bedrock AgentCore Memory
Mem0 Platform + MCP
C 56.1/100Hosted memory layer that extracts facts from conversations and returns the relevant ones for a user, agent or run on later turns.
Verdict An agent can create its own Free account with mem0 init --agent, no email and no card. Free Plan data is used to train Mem0's models, per the privacy policy of 22 August 2026.
Choose it for Chat products that want per-user facts back with one search call and little setup.
Strengths
- An agent can create its own Free account with
mem0 init --agent, no email and no card - Hosted MCP server with 11 tools, listed in the official MCP registry as io.github.mem0ai/mem0
- Public OpenAPI spec, llms.txt and Python and TypeScript SDKs, both released on 2026-09-25
Weaknesses
- Free Plan data is used to train Mem0's models, per the privacy policy of 22 August 2026
- No published rate limits or 429 handling, and only 400 and 404 documented as errors
- Retrieval caps are low below Pro, 1,000 a month free and 5,000 on Starter
Price $19 / moAuth API keyx402 nohosted
Full assessment · Against #1, Amazon Bedrock AgentCore Memory
Cognee
C 54.6/100Open-source memory engine that turns documents, conversations and synced sources into a knowledge graph plus a vector index and answers queries over both.
Verdict Apache-2.0 library, REST server and MCP server, all self-hostable, with local models and no LLM key since 1.6.0. No status page, rate limits, SLA or SOC 2 for Cloud, and Cloud runs in AWS us-east-1 with no EU region.
Choose it for Agents whose memory has to include documents, wikis and chat tools as well as conversation, and for teams happy to self-host.
Strengths
- Apache-2.0 library, REST server and MCP server, all self-hostable, with local models and no LLM key since 1.6.0
- 7-tool MCP server with search_tools and call_tool for reaching the rest on demand
- Public OpenAPI 3.1 file with 46 paths and error models
Weaknesses
- No status page, rate limits, SLA or SOC 2 for Cloud, and Cloud runs in AWS us-east-1 with no EU region
- Cloud calls hung rather than failing when a tenant ran out of credit (August 2026), and the issue is still open
- Billing docs and pricing page disagree on plans and the free allowance
Price $5 / moAuth OAuth or keyx402 nohosted and local
Full assessment · Against #1, Amazon Bedrock AgentCore Memory
Memory (MCP reference server)
C 54.2/100Knowledge-graph persistent memory reference server (entities, relations, observations) stored as JSONL at MEMORY_FILE_PATH.
Verdict Nine tools with typed schemas and accurate read, destructive and idempotent annotations. No pagination or limits. read_graph returns everything and search_nodes every match.
Choose it for A single local agent that wants a small, inspectable store of facts about people and projects.
Strengths
- Nine tools with typed schemas and accurate read, destructive and idempotent annotations
- Plain JSONL storage at a path you choose, easy to back up, diff and edit by hand
- Writes are atomic since 2026.8.31, so an interrupted save can't truncate the file
Weaknesses
- No pagination or limits.
read_graphreturns everything andsearch_nodesevery match - The published version can lose one of two writes made in the same turn. The fix is merged but unreleased
- Search is case-insensitive substring matching, with no ranking or semantics
Price Free · OSSAuth Nonex402 nolocal
Full assessment · Against #1, Amazon Bedrock AgentCore Memory
Disclosure MCP started at Anthropic, which makes the Claude models our research agents and review panel run on (Anthropic donated it to the Agentic AI Foundation, a directed fund under the Linux Foundation, in December 2025), and this server is graded by the same checklist as every other listing.
Graphiti
D 53.3/100Open-source Python framework from Zep that builds a temporal knowledge graph from chat messages, text and JSON.
Verdict Apache-2.0, with FalkorDB, Neo4j or Amazon Neptune as the store. Every add runs LLM extraction, so ingestion costs tokens and time.
Choose it for Teams that want Zep's temporal graph model on their own infrastructure and can run a graph database.
Strengths
- Apache-2.0, with FalkorDB, Neo4j or Amazon Neptune as the store
- 13-tool MCP server over streamable HTTP or stdio, with a Docker Compose file
- Works with OpenAI, Anthropic, Gemini, Groq or a local OpenAI-compatible model
Weaknesses
- Every add runs LLM extraction, so ingestion costs tokens and time
- The MCP HTTP endpoint has no authentication, and no tool carries readOnlyHint or destructiveHint
- Still 0.x, and the last three PyPI releases have no GitHub release notes
Price Free · OSSAuth Nonex402 nolibrary
Full assessment · Against #1, Amazon Bedrock AgentCore Memory
Hindsight
D 50.2/100Memory engine for storing and retrieving information used by agents.
Verdict API keys support bank-level restrictions, expiry and child-key revocation. Retain ingestion costs $10 per million tokens.
Choose it for Agents that should form opinions and summaries from what they stored, such as long-running assistants or copilots that reflect on past sessions.
Strengths
- Keys restricted to named banks, with expiry from an hour to a year and child-key revocation
Memory Defensescreens every retain, with regex redaction even in the open-source server- Published per-token and per-call Cloud prices, with no monthly fee, and a published 99.9 per cent SLA
Weaknesses
- Retain at $10 per million tokens is the priciest ingestion in this category
- 27 MCP tools per bank, with no subset or read-only option
- No status page, no published rate limits and no llms.txt
Price $0.05 / callAuth OAuth or keyx402 nohosted
Full assessment · Against #1, Amazon Bedrock AgentCore Memory
LangMem
D 47.9/100LangMem is LangChain's open-source Python library for long-term agent memory. It extracts facts from conversations with an LLM, stores and searches them in a LangGraph store the owner runs, and includes two memory tools, message summarisation and prompt optimisation.
Verdict Two compact, typed memory tools and a background extraction manager work with any LangGraph store, under the MIT licence with nothing to buy. The newest PyPI release, 0.0.30, dates from 27 October 2025. The repository has no changelog, tags or test workflow, and 23 of 54 open issues have no reply.
Choose it for Teams already on LangGraph that want memory tools and background extraction over a store they run.
Strengths
- MIT licence, installed with
pip install -U langmem, with no account, key or fee of its own - Two agent tools,
manage_memoryandsearch_memory, with typed inputs, an action enum andlimit,offsetandfilteron search actions_permittedlimits the manage tool to any subset of create, update and delete, andcreate_memory_store_managerleaves deletes off by default
Weaknesses
- The newest PyPI release, 0.0.30, is from 27 October 2025, and every commit to
srcon main is older than that - No changelog, GitHub releases or tags, and versions are still 0.0.x
- The repository's two workflows deploy docs and publish to PyPI. Neither runs the tests
Price Free · OSSAuth Nonex402 nolocal
Full assessment · Against #1, Amazon Bedrock AgentCore Memory
Head to head
- Amazon Bedrock AgentCore Memory vs Zep A 79.4 vs B 69.3
- Amazon Bedrock AgentCore Memory vs Honcho A 79.4 vs B 64.1
- Amazon Bedrock AgentCore Memory vs Supermemory API + MCP A 79.4 vs B 63.6
- Amazon Bedrock AgentCore Memory vs Mem0 Platform + MCP A 79.4 vs C 56.1
- Honcho vs Zep B 64.1 vs B 69.3
- Supermemory API + MCP vs Zep B 63.6 vs B 69.3
- Mem0 Platform + MCP vs Zep C 56.1 vs B 69.3
- Honcho vs Supermemory API + MCP B 64.1 vs B 63.6
- Honcho vs Mem0 Platform + MCP B 64.1 vs C 56.1
- Mem0 Platform + MCP vs Supermemory API + MCP C 56.1 vs B 63.6
Questions
What are the highest-rated memory layers for AI agents?
Amazon Bedrock AgentCore Memory has the highest benchmark score of the 10 ranked memory layers, 79.4 (A). Zep is second with 69.3 (B).
How many memory layers are agent-ready?
1 of the 10 ranked here grade BB or better, the bar for agent-ready on the Anchor benchmark.
Which memory layers accept x402 payments?
Honcho. An agent can pay these per call in USDC with no account.
How is this list ranked?
By the Anchor benchmark score out of 100, a weighted mean of the scored categories minus deductions for negative events, from public evidence re-checked as vendors change. Listings cannot pay for a place. The latest assessment behind this page is from 8 October 2026.
How this list is made
The order is the Anchor benchmark score, the same number as on each listing and in the top list. Each listing is graded from public evidence against the benchmark checklist, and the picks above are worked out from those grades, prices and facts. No listing pays for its place, and paid audits or listing help never change a score.
Full ranked table · 55 head-to-head comparisons · Best tools in every category