Automem
AutoMem gives MCP-based AI agents persistent graph-vector memory, storing each memory in FalkorDB for relationships and Qdrant for meaning so recall returns
AutoMem is the memory layer we'd hand a team already living inside MCP clients: hybrid graph plus vector recall, one-command install, and MIT licensing so you can run it yourself. The numbers are worth taking seriously precisely because the vendor also publishes the loss — 57.4% BEAM at 10M tokens and roughly 2.6–4.8k context tokens per answer against a board leader feeding 17–27k. It is second to Hindsight on that axis and ahead of Honcho at every tier. Go managed on Railway instead if nobody on the team wants to touch a Dockerfile.
Verified 9d ago · liveness 74/100 · cite: rightaichoice.com/tools/automem
- Developers building MCP-based agents that need relational memory across sessions
- Teams running Claude Code, Cursor or Codex CLI who want shared project context
- Privacy-first orgs that need memory kept in their own VPC or on local Docker
- Engineers optimizing context cost who want flat per-answer token spend at scale
- Non-technical buyers who need memory enabled entirely from a dashboard
- Teams whose agent stack does not speak MCP
- Projects wanting to hand-tune recall weights — 0.16 built four knobs and deliberately left them off
We scan live Reddit threads, YouTube comments, X posts, G2 reviews and other communities — and hand you an honest verdict in under a minute.
- Honest verdict, not marketing
- Real pros & cons from real users
- Attributed quotes with receipts
3 free scans · no card needed
Skip AutoMem if your agent stack does not speak MCP, or if you want recall weights you can hand-tune — 0.16 built four tuning knobs and deliberately left them off for correctness.
Self-hosting is free under MIT, but you pay for the FalkorDB and Qdrant instances you run to hold the graph and vector index.
Deployment economics split three ways, and only two of them are metered. Local Docker self-hosting is free under the MIT licence — you pay only for the FalkorDB and Qdrant instances you run. Railway managed cloud is pay-as-you-go, which suits a small team that wants memory live today without owning infrastructure. InstaPods is for orgs that already run Kubernetes in their own VPC and would rather absorb the cost internally than send memory to a hosted service. Against managed memory vendors,
In short
Automem — AutoMem gives MCP-based AI agents persistent graph-vector memory, storing each memory in FalkorDB for relationships and Qdrant for meaning so recall returns. Best for Developers building MCP-based agents that need relational memory across sessions, Teams running Claude Code, Cursor or Codex CLI who want shared project context, Privacy-first orgs that need memory kept in their own VPC or on local Docker. Free to use.
What's new in Automem
Checked 9 days agoAcross the latest 3 updates: 1 launch and 2 news mentions.
AutoMem on the Neutral Benchmark: The Numbers That Hold Up
AutoMem reports graceful BEAM scaling to 10M tokens and roughly 3.9k context tokens per answer on the neutral Agent Memory Benchmark, and states the conversational-recall gap plainly on the same page as the win.
AutoMem 0.16: Correctness Over Knobs
Version 0.16 ships with four recall-tuning knobs built but deliberately left off, choosing correctness over configurability; the vendor notes its internal benchmark numbers are not yet comparable to other systems'.
Agent Memory in 2026: An Honest Comparison of Mem0, Zep, Letta, and the Rest
AutoMem publishes a comparison of agent memory systems covering Mem0, Zep and Letta.
What people actually say about Automem — is it worth it?
We ran a structured research pass across product reviews, community discussions, and post-purchase forum threads to surface the patterns vendors won't publish themselves. Below: the recurring strengths, the hidden costs people mention most, and the cohort that consistently regrets adopting this tool.
11 mentions across 2 sources (Hacker News, Lemmy) · researched Jul 3, 2026.
Average across the 2 sources that answered — each source counts once, not each post.
- +Open-source MIT license allows full customization and self-hosting.
- +Hybrid graph-vector retrieval offers richer memory than simple key-value.
- +One-command curl install gets you running quickly.
- +Integrates with any MCP-compatible client like Claude Code or Cursor.
- +Background consolidation improves recall quality over time.
- −Docker requirement blocks users without Docker or on restricted systems.
- −Documented real-world deployments are scarce — early adopter risk.
- −Competing with Claude's built-in auto-memory which is simpler.
- −No native memory retention without active MCP client integration.
- −Graph-vector complexity may be overkill for simple memory needs.
- • Cloud hosting on Railway incurs usage costs; no clear free tier.
- • Self-hosting requires compute resources for FalkorDB and Qdrant.
Viability Score
How well maintained and how widely used is Automem? Built from what the vendor actually publishes (docs, changelog, tutorials, integrations, pricing), whether the site is live, and how much real users discuss it. How we calculate this
Last calculated: October 2026
How we score →Key Features
- Hybrid graph-vector memory using FalkorDB for relationships and Qdrant for meaning
- Single hybrid recall request combining graph queries, vector search, keyword matching and temporal signals
- 11 authorable relationship types (LEADS_TO, CONTRADICTS, EXEMPLIFIES and others) plus system-generated semantic and temporal edges
- Background consolidation that clusters, strengthens and decays memories
- Automatic entity extraction and pattern detection in the enrichment worker
- Graceful degradation to graph-only mode when Qdrant is unavailable
- Remote MCP over SSE and Streamable HTTP for mobile and desktop clients
- MCP bridge package @verygoodplugins/mcp-automem translating MCP calls into HTTP API requests
- Single-command install via curl -fsSL get.automem.ai | sh or npx @verygoodplugins/mcp-automem setup
- Deploy locally on Docker, managed on Railway, or self-hosted on InstaPods
- 3D memory graph visualization with hand gesture controls and 11 relationship types
- Real-time monitoring of the memory nexus and agent graph
- Batch embedding generation via a dedicated EmbeddingWorker
- ConsolidationScheduler with decay, creative, cluster and forget cycles
- SyncWorker for drift repair between memory stores
About Automem
AutoMem is an open-source, MIT-licensed memory layer for AI agents that stores every memory twice: in a FalkorDB knowledge graph for relationships and time, and in a Qdrant vector index for meaning and similarity. The point is that recall returns more than the closest-sounding paragraph — it returns the entity, the project it belongs to, and the decision you settled weeks ago. Developers wiring memory into Claude Code, Cursor, Codex CLI, ChatGPT's code interpreter, Claude Desktop, Windsurf, GitHub Copilot or any other MCP-compatible client get a single endpoint that all of them share, because the mcp-automem bridge package (@verygoodplugins/mcp-automem) translates MCP calls into HTTP requests against one AutoMem service. Deployment comes in three shapes with the same MCP surface: local Docker for full control and offline work, Railway for a one-command managed cloud, and InstaPods for self-hosted cloud inside your own VPC. Setup is a single curl or npm command that detects the environment, provisions AutoMem and wires up your MCP clients. Remote MCP over SSE and Streamable HTTP means mobile clients such as the ChatGPT and Claude apps can connect too, not only desktop CLIs. What separates it from plain vector memory is the background consolidation layer: new memories get clustered, links you keep returning to get strengthened, noise decays, and the graph gets richer than the one you stored. Version 0.10.0 added a 3D memory graph visualization with 11 relationship types, hand gesture controls and real-time monitoring of the memory nexus, while 0.16 deliberately ships without four recall-tuning knobs built but left off in favour of correctness. On the neutral Agent Memory Benchmark harness, AutoMem holds 57.4% BEAM accuracy at 10M tokens while feeding the answerer roughly 2.6–4.8k context tokens per answer — flat at scale, where the board leader feeds 17–27k. It is second to Hindsight on that axis and ahead of Honcho at every tier. The vendor also publishes its losses: the conversational-recall gap behind the leader is on the same page as the win.
Behind the Verdict
AutoMem's bet is that vector-only memory is a half-answer. Similarity search finds text that sounds like your question; it does not know that this project uses that tool, that this decision contradicts last month's, or that the fix you settled on Thursday came from an install-script error on WSL. AutoMem stores each memory in FalkorDB as an entity with typed relationships — 11 authorable types including LEADS_TO, CONTRADICTS and EXEMPLIFIES, plus system-generated semantic and temporal edges — and in Qdrant for semantic match. A single hybrid recall request runs a graph query and a vector search together, so you get the answer plus the thread it belongs to. The architecture is genuinely separated rather than a wrapper: the MCP bridge is a Node.js package (@verygoodplugins/mcp-automem) that translates protocol calls into HTTP requests to a Flask API service (Gunicorn on :8001), which owns the dual storage, background workers for enrichment, embedding generation, consolidation and drift repair. Multiple AI clients share one memory store as a result, and the service stays platform-independent. The honest caveat is the graceful-degradation design: if Qdrant is unavailable the system continues in graph-only mode, which keeps you running but is a different recall quality than the hybrid path. Strength one: cost behaviour. The benchmark page reports ~2.6–4.8k context tokens per answer flat at every scale, against 17–27k for the board leader. If you are paying per token on every agent turn, a flat context budget is the difference between a demo and a product. Strength two: consolidation. New memories get clustered, returning links get strengthened, noise decays. The graph you read tomorrow is not the graph you wrote today. Strength three: three deployment shapes with one MCP endpoint, including InstaPods for running inside your own VPC on Kubernetes — which is what a privacy-first org actually needs to pass review. Where it does not fit: this is not a dashboard product. There is no point-and-click way to enable memory for a non-technical buyer; you run a curl or npx command, choose local or cloud, and wire MCP clients. If your stack does not speak MCP, AutoMem is not for you. And if you want to hand-tune recall weights, version 0.16 built four tuning knobs and deliberately left them off — that is the vendor choosing correctness over configurability, and if you disagree you are on the wrong tool. Finally, take the internal numbers with the caveat the vendor prints itself: 87.00% LongMemEval full with 97.00% recall@5 and 84.74% on LoCoMo full are internal engineering baselines, not cross-system comparisons. The neutral harness is where the comparable 57.4% BEAM figure lives, and the conversational-recall gap is real and stated. Read it before you buy the pitch.
Researching Automem? Get your full AI stack in 60 seconds.
Free, no signup — tell us your goal and get tools matched to your budget & existing stack.
Real-world workflow fit
Concrete scenarios for the personas Automem actually fits — and what changes day-one when you adopt it.
Runs curl -fsSL get.automem.ai | sh, picks local Docker, and AutoMem detects the environment, provisions itself and wires the Claude Code MCP client. Project conventions, install-script fixes and decisions land in the graph as work happens.
Outcome: Next session opens with the conventions and last week's decisions already recalled instead of re-explained, and context spend stays in the couple-thousand-token range per answer rather than growing with history.
Both point at the same Railway-hosted AutoMem endpoint over remote MCP, so the Cursor editor and the ChatGPT code-interpreter session read and write one shared memory store instead of two private ones.
Outcome: A decision made in Cursor is available in the next ChatGPT session, and no one has to paste context between tools.
Deploys AutoMem through InstaPods into the company VPC on Kubernetes, keeps FalkorDB and Qdrant inside the same perimeter, and reviews the architecture docs covering the MCP bridge and dual-storage layers.
Outcome: Agents get cross-session memory without memory leaving the company's infrastructure, which is what the security review needed to clear it.
Use Cases
- Give Claude Code persistent recall of project conventions and decisions you settled in earlier sessions
- Build a support agent that recalls each customer's previous conversations without re-prompting the history
- Run one shared memory store across Cursor, Codex CLI and ChatGPT so context follows you between tools
- Log and retrieve debugging sessions, including the WSL install-script errors you already fixed once
- Keep a team knowledge base that MCP agents read from and write to as work happens
- Keep agent context spend flat as conversations grow past 10M tokens
- Run agent memory inside your own VPC on Kubernetes for privacy or compliance review
- Connect ChatGPT or Claude mobile apps to the same memory your desktop CLI writes to
Models Under the Hood
as of 2026-10-10
Limitations
- AutoMem is an open-source hybrid graph-vector memory layer for AI agents, accessed through MCP rather than a dashboard.
- Version 0.16 deliberately left off four recall-tuning knobs that were built, choosing correctness over configurability, so this is not the tool for teams that want to hand-tune recall weights.
- The vendor states a conversational-recall gap in its own benchmark write-up, which means pure chat recall is not where AutoMem leads against alternatives, and it notes its internal benchmark numbers are not yet comparable to other systems'.
- If Qdrant is unavailable the system falls back to graph-only mode, which keeps operations running but is a different recall path than the hybrid one.
as of 2026-10-02
Verification history
We have re-verified Automem 6 times since . Each pass re-reads the vendor's own pages and re-checks every listed field against that evidence; passes where nothing had changed are marked as such.
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
Free to cite with attribution — this page re-verifies continuously.
12-month cost
Project the real annual outlay, including the implied monthly cost when only an annual tier is published.
Vendor list price only. Add-on usage, seat overages, and contract minimums are surfaced under Hidden costs & gotchas.
Plans compared
For each published Automem tier: who it actually fits, and what it adds vs. the previous tier. Cross-reference the cost calculator above for projected annual outlay.
Local (Self-Hosted)
$0/mo
Ideal for
A developer or small team already running Docker who wants agent memory to stay on their own machine and keep working offline.
What this tier adds
Starting tier and the free entry point: MIT-licensed source, everything runs locally, no external dependencies, same MCP endpoint as the cloud options.
Railway (Managed Cloud)
Pay-as-you-go
Ideal for
A team that wants memory live today and would rather not own infrastructure — including users who need a remote MCP endpoint for the ChatGPT and Claude mobile apps.
What this tier adds
Adds one-command cloud deploy with automatic updates and backups and global availability, replacing the DIY hosting you handle on the local tier.
InstaPods (Self-Hosted Cloud)
Free tier available; paid for larger graphs
Ideal for
Privacy-first and enterprise teams that already run Kubernetes in their own VPC and must keep agent memory inside their own infrastructure.
What this tier adds
Moves the managed cloud experience inside your own VPC, Kubernetes-native, with data staying in your infrastructure rather than a hosted service.
Where the pricing makes sense
The company stage and team size where Automem's pricing actually pencils out — and where peers do it cheaper.
Deployment economics split three ways, and only two of them are metered. Local Docker self-hosting is free under the MIT licence — you pay only for the FalkorDB and Qdrant instances you run. Railway managed cloud is pay-as-you-go, which suits a small team that wants memory live today without owning infrastructure. InstaPods is for orgs that already run Kubernetes in their own VPC and would rather absorb the cost internally than send memory to a hosted service. Against managed memory vendors,
Setup time & first value
How long it actually takes to get something useful out of Automem — broken out by persona, not the marketing-page minute.
Solo developer on local Docker: minutes to first recall — one curl or npx command detects the environment, provisions AutoMem and wires your MCP client. Team on Railway managed cloud: same single command with the cloud install target, and updates and backups are handled for you. Platform engineer on InstaPods: longer, since it runs in your own VPC on Kubernetes and needs your cluster capacity and
Switching to or from Automem
How to bring data in from common predecessors and how to get it back out — written for the switcher, not the buyer.
- →From pasting context into each prompt: point your MCP client at AutoMem and let project conventions and decisions accumulate in the graph instead of a scratch file.
- →From Mem0: AutoMem publishes an honest comparison of Mem0, Zep and Letta, so read that page before you rip anything out and map your existing memories onto AutoMem's store and recall operations.
- →From a vector-only memory store: keep the semantic half, add the FalkorDB graph so recall can return the entity and its typed relationships rather than only the closest paragraph.
- ↗To Mem0 or Zep: both are covered in AutoMem's own agent-memory comparison, so start with that page and plan how your graph relationships map to their memory model.
- ↗To Hindsight: if BEAM accuracy at 10M tokens is the deciding axis, Hindsight leads it — check its deployment shapes against your Docker, Railway or VPC setup first.
- ↗To Honcho: only worth the move if you need something AutoMem does not cover; on the neutral BEAM axis AutoMem is ahead of Honcho at every tier.
Integrations
Resources & Guides
- Documentationautomem.ai
Overview · Automem
Full product docs from automem.ai
- Documentationautomem.ai
Docs · Automem
Full product docs from automem.ai
- Resourceautomem.ai
Skill · Automem
Helpful link from automem.ai
- Resourceautomem.ai
Benchmarks · Automem
Helpful link from automem.ai
- Resourceautomem.ai
Blog · Automem
Helpful link from automem.ai
Tutorials & Learning
YouTube returned 6 videos for “Automem”, and we withheld 6: 6 could not be judged, because “Automem” is a single word that other videos use for other things. We are showing none, because we could not prove any of them are about Automem.
Official links
Tools that pair well with Automem
Common stack mates teams adopt alongside Automem, with the specific reason each pairing earns its keep.
Cognee
Open-source memory platform giving AI agents graph-based, relationship-aware recall with citations.
Powermem
Open-source, self-hosted memory layer that gives AI agents persistent recall via hybrid retrieval.
EverMemOS
Memory operating system that gives AI agents persistent, self-evolving recall across sessions, platforms, and teams.
Featured Head-to-Head Comparisons
Automem vs Spider Cloud
If you need persistent, relational memory for your AI agent, Automem is the clear winner; it’s open source and integrates directly with Claude and Cursor. If instead you need reliable web scraping for RAG, Spider Cloud’s Rust engine and AI extraction offer unbeatable speed and cost efficiency. They solve different problems—choose based on whether your bottleneck is memory or data access.
Automem vs Presto Voice
Automem and Presto Voice serve completely different markets. Automem is a developer tool for AI agent memory, perfect for teams building persistent-context assistants. Presto Voice is a voice AI platform for QSR drive-thrus focused on revenue uplift. Choose based on your domain: agent memory vs. restaurant automation.
Automem vs Temporal Ai
Choose Automem if you need persistent relational memory for AI agents across chat sessions, especially with MCP-compatible tools. Choose Temporal if you need reliable orchestration of long-running workflows with crash recovery and human-in-the-loop. They solve different problems: memory vs execution.
Alternatives to Automem
View all
Frequently Asked Questions
Used Automem? Help shape our editorial sentiment research.