Mem0 vs Zep vs Upstash Redis: Real AI Agent Memory Cost at Scale in 2026
The short answer: these three are not substitutes for each other, and the price gap between them reflects that. Mem0 and Zep are managed memory-extraction platforms — they read a conversation, decide what's worth remembering, and store it in a structured, retrievable form — priced from $19 to $249+/month on Mem0 and $125 to $312+/month on Zep at their published self-serve tiers. Upstash Redis is raw key-value and vector storage with no extraction intelligence built in, priced at roughly $0.20 per 100,000 commands on its pay-as-you-go tier, which is dramatically cheaper per operation but requires a team to build its own memory-extraction logic on top. At a moderate agent workload (2 million memory writes and 5 million retrievals a month), that difference plays out as $249 or more on Mem0's top self-serve tier (Pro), likely requiring a custom Enterprise quote beyond its published 500,000-memory allowance, against about $14 a month in raw Redis command costs on Upstash for storing and retrieving the same volume of records — the Upstash figure buys storage only, not the extraction and summarization work Mem0 and Zep are built to do.
What each vendor bills, and what it actually does
Mem0 (OFFICIAL, Mem0 pricing page, current after a 2026 free-tier increase). A managed memory-extraction API: ingests conversation turns, extracts facts worth remembering, and retrieves relevant memories scoped by user, agent, or session ID. Free (Hobby): 10,000 memory adds and 1,000 retrievals a month — tripled from a 1,000-memory allowance earlier in 2026, now enough for real prototyping rather than just a demo. Starter: $19/month, 50,000 memories. Growth: $79/month. Pro: $249–250/month, 500,000 memories, and the tier that unlocks graph memory (structured relationship extraction, avoiding the need to self-host a separate graph database). Enterprise: custom, adding on-premises deployment, SSO, and audit logging. No public overage or top-up rate beyond Pro's 500,000-memory allowance was found in the sources reviewed; a workload needing more requires a custom Enterprise quote.
Zep (OFFICIAL for the mechanism, REPORTED for exact current rates). A temporal knowledge-graph memory layer built on Graphiti (Zep's open-source, Apache-2.0-licensed graph engine, which remains freely self-hostable on Neo4j or FalkorDB). Zep restructured its pricing in 2026, moving away from pure pay-as-you-go and retiring its self-hosted Community Edition (the open-source Graphiti engine itself is unaffected). Free: very limited, reported around 1,000 credits a month, described as exhaustible within days of active development. Paid tiers now bill in credits, starting at Flex around $125/month, with a top listed self-serve rate around $312/month in some sources; the exact credit-to-operation conversion was not confirmed in the sources reviewed. Zep's differentiator is tracking how facts change over time, not just what facts exist — relevant for CRM-style agents or long-running projects where a user's preferences or project state evolve.
Upstash Redis (OFFICIAL, Upstash pricing page). Raw, serverless Redis-compatible storage with no memory-extraction logic of its own — a team using Upstash as a memory backend must build the extraction, summarization, and retrieval logic itself. Free: 256 MB storage, 500,000 commands/month, 10,000 vector documents. Pay-as-you-go: $0.20 per 100,000 commands, storage billed separately, up to 100 GB. Fixed plans (unlimited commands within a provisioned tier): $10 to $1,500/month, useful once pay-as-you-go's per-command cost would exceed a fixed tier's flat rate at high, steady traffic. At 100 million commands a month, pay-as-you-go alone was reported at about $200.
Normalizing the workload
All ILLUSTRATIVE. This article models a moderate production agent memory workload: 100,000 monthly active users, each generating about 20 memory-write operations and 50 memory-retrieval operations a month — 2 million writes and 5 million retrievals in total, 7 million operations combined.
Cost at this workload
Formula: Mem0/Zep = whichever published tier's memory allowance covers the volume, or QUOTE-ONLY / UNKNOWN if it exceeds every published self-serve tier; Upstash = total operations ÷ 100,000 × $0.20, treating each memory write or retrieval as one Redis command (an ILLUSTRATIVE simplification; a real implementation may use more than one command per logical memory operation).
| Platform | Basis | Cost at 2M writes / 5M retrievals a month |
|---|---|---|
| Mem0 | 2 million memories exceeds Pro's 500,000-memory allowance by 4x; no public overage rate | QUOTE-ONLY / UNKNOWN beyond Pro ($249–250/month), likely requires an Enterprise quote |
| Zep | Credit-to-operation conversion not confirmed; workload likely exceeds the ~$312/month top listed tier | QUOTE-ONLY / UNKNOWN, likely requires an Enterprise quote |
| Upstash Redis, pay-as-you-go | 7,000,000 operations ÷ 100,000 × $0.20 | $14.00 (storage billed separately) |
The two purpose-built memory platforms cannot be priced with confidence at this volume from public information alone — both are structured around allowances that top out well below this article's illustrative workload, with Enterprise sales conversations standing in for a public rate card past that point. Upstash's figure is a real, confirmed number, but it buys only the storage substrate, not the memory-extraction intelligence Mem0 and Zep are built around.
What the Upstash figure doesn't include
Using Upstash Redis as a memory backend means building, and maintaining, the logic that decides what's worth remembering, extracts it from raw conversation text, resolves conflicts when a new fact contradicts an old one, and structures retrieval so the right memories surface at the right time. This is real, ongoing engineering work with its own cost — Mem0's and Zep's subscription fees are, in large part, a purchase of that already-built logic, not merely storage. A team choosing Upstash for cost reasons should budget engineering time for this layer, not just the near-zero Redis command cost.
Where each purpose-built platform's ceiling actually sits
Formula: Mem0's confirmed self-serve ceiling = 500,000 memories/month (Pro); Zep's confirmed self-serve ceiling ≈ $312/month worth of credits, with the exact operation-count equivalent unconfirmed. A workload of 2 million memory writes a month is 4 times Mem0's Pro allowance and, on the same illustrative operation-count basis, plausibly well beyond Zep's top listed credit tier as well — both platforms are structured for a self-serve ceiling meaningfully below this article's moderate-scale example, making an Enterprise conversation the realistic path for any agent product with a genuinely large or fast-growing user base.
Sensitivity
- Memories per user per month. The single largest lever for both Mem0 and Zep; a lighter-touch agent (fewer facts extracted per session) could fit within Pro's or Flex's allowance at meaningfully higher user counts than this article's illustrative 20-writes-per-user assumption.
- Commands per logical memory operation on Upstash. A real implementation likely issues more than one Redis command per memory write or retrieval (an index update, a TTL set, a vector similarity search); this article's 1:1 assumption is a floor, not a ceiling.
- Graph memory requirement. Needing Mem0's graph-memory feature forces the Pro tier regardless of raw memory volume, since it is a feature gate, not a volume-based unlock.
- Fixed-plan crossover on Upstash. At sufficiently high and steady command volume, one of Upstash's fixed plans ($10–$1,500/month unlimited commands) becomes cheaper than continuing on pay-as-you-go.
Budgeting traps
- Assuming Mem0's or Zep's published top tier scales linearly to production volume. Both cap out at a specific self-serve allowance, past which the real cost is a custom quote, not an extrapolated per-unit rate.
- Comparing Upstash's raw command cost directly to Mem0's or Zep's subscription price as if they buy the same thing. One buys storage; the other buys extraction and retrieval intelligence built on top of storage.
- Underestimating the engineering cost of building memory logic on Upstash. The near-zero Redis command cost does not include the ongoing work of deciding what to remember and how to retrieve it well.
- Missing that Zep retired self-hosted Community Edition. Any plan built around free self-hosted Zep specifically (as opposed to the still-open-source Graphiti engine underneath it) needs to be re-evaluated against Zep's current managed-only paid tiers.
What to ask before you buy
Ask Mem0 and Zep directly for a quote at your actual expected memory-write and retrieval volume once it exceeds their published self-serve ceilings, since neither publishes a confirmed overage rate past their top listed tier. If evaluating Upstash as a build-your-own memory backend, budget separately and explicitly for the engineering time to build extraction and retrieval logic, since the Redis command cost alone understates the real cost of that path.