Memory Layer

Vital signs

Why it's on the table

On the table, Memory Layer (Mm) is seat 16 of 58, in the Knowledge & Memory family. It is an experimental element — promising, volatile, and worth a contained experiment rather than a commitment. Budget curiosity, not dependence. It is optional: plenty of companies run without it — until a specific trigger (scale, regulation, cost, or customers) makes it essential for them. It sits in the lowest paid band — lunch money against the hours it returns.

Memory Layer: the top 5 — v2026.Q3

Edition v2026.Q3 · ranking, pricing and status verified 2026-08-06.

1. 1Mem0Mem0 (YC S24)

   Hobby free (10k adds/mo) · Starter $19/mo · Pro $249/mo · Enterprise custom (on-prem, SSO)

   Best for Adding 'remembers the user' to an existing app in an afternoon — a hosted extraction-and-retrieval API that works with any model and framework.

   The category's center of gravity: 61.7k GitHub stars, 90k+ registered developers, 13M+ package downloads, and API calls that grew 35M (Q1 2025) to 186M (Q3 2025). Raised $24M (seed + Series A led by Basis Set, Oct 2025) and became the exclusive memory provider for AWS's Agent SDK. Still shipping fast: Dream, its background memory-consolidation engine, launched Aug 4, 2026.

   Watch Its 'SOTA' LoCoMo claims are disputed — Zep published a rebuttal showing a corrected implementation beat Mem0's setup, and Mem0's own paper had a full-context baseline outperforming it. Retrieval quotas bite: 1k retrievals/mo free and 5k on the $19 tier push retrieval-heavy apps to $249 quickly.

   $24M raised (seed $3.9M + $20M Series A led by Basis Set); API calls 35M Q1 → 186M Q3 2025 (Oct 28, 2025) [src] · 61.7k GitHub stars, Apache-2.0; new memory algorithm shipped Apr 2026 (checked Aug 2026) [src] · Dream background memory consolidation launched Aug 4, 2026; 90k+ developers on platform [src]
   Site →Docs →

2. 2ZepZep AI

   Free 10k credits/mo · Flex $1,250/yr · Flex Plus $3,750/yr · Emerging Cos $13k/first year (SOC 2, HIPAA BAA) · Enterprise custom

   Best for Enterprises whose agents must remember evolving business facts — temporal knowledge graphs over chat plus CRM/app data, with provenance tracking and access control.

   The engineering-serious enterprise pick: temporal context graphs (open-sourced as Graphiti, 29k stars) with sub-200ms retrieval at 100M graphs, provenance lineage for synthesized facts (Jul 2026), attribute-based access control (Jul 2026), and SSO-gated memory over MCP (Jun 2026). Customers include Zscaler, Samsung, and HoneyBook; S&P Global called it a likely 'de facto partner in this layer of the enterprise agent stack.'

   Watch No cheap paid tier — the jump from free to $1,250/yr excludes hobbyists, and the self-hosted Community Edition was deprecated (code moved to legacy/), so the real product is cloud-only. Its 94.7% LoCoMo / 90.2% LongMemEval numbers are vendor-run, like everyone else's.

   94.7% LoCoMo at 155ms, 90.2% LongMemEval at 162ms; 161–168ms retrieval at 10M–100M graphs (vendor, Aug 2026) [src] · Graphiti OSS: 29k stars, v0.29.2 released Jun 8, 2026 [src] · Provenance tracking (Jul 14, 2026), ABAC (Jul 9, 2026), enterprise-SSO memory MCP server (Jun 30, 2026) [src]
   Site →Docs →

3. 3LettaLetta (ex-MemGPT, UC Berkeley)

   Free · Pro $20/mo (20 stateful agents) · Developer $0.10/active agent/mo + $0.00015/sec tool exec · Enterprise custom

   Best for Building agents whose memory and self-improvement are the product — stateful agents that learn across sessions, from the researchers who invented the pattern.

   The intellectual origin of the category: MemGPT (Oct 2023) invented LLM virtual context management, and Letta commercialized it with a $10M Felicis-led seed at $70M (Sep 2024). Its March 2026 pivot doubled down on what worked — Letta Code, a model-agnostic harness with git-backed memory files ('you own the memory, you choose the model'), plus research on sleep-time compute and continual learning that the rest of the field imitates.

   Watch Strategic churn is real: the Mar 16, 2026 'next phase' deprecated large chunks of the platform (Filesystem, server-side templates, MCP integrations, sleep-time agents, tool rules), and the original server repo is now labeled legacy. Smallest disclosed war chest of the leaders.

   $10M seed led by Felicis at $70M post (Sep 23, 2024); founders created MemGPT at Berkeley's Sky Lab [src] · Pivot to Letta Code harness with git-backed memory announced Mar 16, 2026; legacy features deprecated by mid-April [src] · 24k stars (Apache-2.0) on the letta repo, now labeled the legacy V1 API server (Aug 2026) [src]
   Site →Docs →

4. 4SupermemorySupermemory

   Free (~$5 usage) · Pro $19/mo · Max $100/mo · Scale $399/mo · Enterprise custom (self-host)

   Best for High-volume, cost-sensitive context: memory + RAG + connectors (Slack, Gmail, Drive, GitHub) in one API, plus a personal memory app that follows you across AI tools.

   The fastest riser: 1.5B+ memories stored, sub-300ms recall claims, and a shipping pace that produced a POSIX-compatible semantic filesystem (SMFS, May 28, 2026 — claimed 55% cheaper agentic retrieval), Context Cloud (May 18, 2026), and default 'dynamic dreaming' consolidation (May 25, 2026) — all on a $3M pre-seed (Susa Ventures, Oct 6, 2025). Vendor cites internal deployments at Google and Nissan.

   Watch Tiny funding versus rivals and a split focus (consumer app + infra API). Benchmark leadership claims (LongMemEval, LoCoMo, ConvoMem) and the Google/Nissan logos are vendor-sourced with no independent verification.

   1.5B+ memories saved; sub-300ms recall claimed (vendor, Aug 2026) [src] · $3M pre-seed led by Susa Ventures, with Browder Capital and SF1 (Oct 6, 2025) [src] · SMFS semantic filesystem launched May 28, 2026 claiming 55% cheaper agentic retrieval [src]
   Site →Docs →

5. 5CogneeTopoteretes

   OSS free (Apache-2.0) · Cloud: Free 1M tokens · Standard $2.50/1M tokens + $5/workspace · Enterprise custom (BYO cloud)

   Best for Teams who want to own the memory pipeline — an open-source ECL (extract-cognify-load) engine that builds combined knowledge-graph + vector memory over your own databases.

   The credible open-source alternative to hosted memory APIs: 29.8k stars, 5M+ SDK runs monthly, v1 shipped, and real production proof (Bayer runs agentic research memory on it; Knowunity POC'd 40,000 students in 2 days). Deploys self-hosted, Docker, on-prem, or cloud, and plugs into Claude Code, Cursor, LangGraph, and MCP.

   Watch The smallest commercial operation in the top 5 — funding undisclosed (Pebblebed and others, amounts unannounced) — and the ECL pipeline demands data-engineering appetite that Mem0's two-line SDK doesn't.

   29.8k GitHub stars, 5M+ SDK runs monthly, v1 released (vendor, Aug 2026) [src] · Bayer production deployment of agentic research memory; Knowunity 40k-student POC in 2 days (vendor case studies) [src] · Cloud pricing $2.50/1M tokens + $5/workspace, free 1M-token tier (Aug 2026) [src]
   Site →Docs →

Memory Layer: the top 8 compared

Edition v2026.Q3 · ranking, pricing and status verified 2026-08-06.

How to choose your memory layer

If you have a working app and just need it to remember users across sessions, this week

Mem0 — free for 10k memory-adds a month, $19 after, two-line SDK, works with any model. Watch the retrieval quota, not the add quota.

If you're an enterprise whose agents must track facts that change over time — accounts, policies, patient state — with audit and access control

Zep. Temporal graph with provenance and ABAC is the point; budget $1,250/yr minimum, $13k if you need SOC 2 II + HIPAA BAA as a startup.

If the agent itself is the product and it must demonstrably learn and improve over weeks

Letta — the MemGPT lineage, sleep-time compute, and git-backed memory you can inspect and version. Accept the platform-pivot risk.

If you're ingesting everything — email, Slack, docs, screenshots — and cost per token retrieved decides the architecture

Supermemory: connector-heavy, sub-300ms claims, and SMFS cut agentic retrieval costs 55% by the vendor's own math. Verify their benchmarks against your data.

If one assistant, one user, one vendor — 'my chatbot should remember me'

Don't buy a layer. Claude and ChatGPT memory do this natively, and Anthropic's client-side memory tool is GA for developers. A standalone memory layer only earns its keep across models, apps, or agents.

Memory Layer: the whole field

19 more tools tracked in this category, including 6 dead, renamed, or sunsetting — a reference that hides the graveyard isn't one. Verified 2026-08-06.

Memory Layer: the category in numbers

Edition v2026.Q3 · ranking, pricing and status verified 2026-08-06.

• Mem0 raised $24M (Oct 28, 2025) on API-call growth from 35M (Q1 2025) to 186M (Q3 2025) and became AWS Agent SDK's exclusive memory provider [src]
• Bundling from above: AWS AgentCore Memory meters memory at $0.25/1k events, Google ships Vertex Memory Bank, Redis launched Iris, and Anthropic's developer memory tool went GA — the primitive is commoditizing [src]
• Native assistant memory became table stakes: Claude memory launched Oct 23, 2025 (559 HN points); ChatGPT moved to fully automatic, continuously-updated memory [src]
• 2026's feature battleground is consolidation: Supermemory made 'dynamic dreaming' default (May 25), Mem0 shipped Dream (Aug 4), Letta published sleep-time compute research — everyone now sleeps [src]
• Benchmark credibility crisis: Zep's rebuttal (May 2025, updated Jun 2026) showed Mem0's LoCoMo comparison used a flawed Zep implementation, and Mem0's own paper had a full-context baseline (~73%) beating its system (~68%) [src]
• Architecture still unsettled: Letta's Mar 16, 2026 pivot moved memory from server-side databases to git-backed files and deprecated much of its platform API — the field's founder rethinking the field's premise [src]

Memory Layer: method & sources

Ranking charter: ecosystem adoption, verified traction, shipping velocity, and pricing accessibility — explicitly NOT vendor benchmark scores, because every vendor here claims to lead LoCoMo/LongMemEval and the Zep-Mem0 dispute plus LoCoMo's full-context-baseline problem make those numbers unusable for ranking. Conflicts resolved: Supermemory's raise is reported as $2.6M in some coverage and $3M on the vendor blog — we cite the vendor's own Oct 6, 2025 post ($3M, Susa-led); single-source, flagged. Mem0's site says 62,590 stars while the GitHub page showed 61.7k the same day — timing/rounding, we cite GitHub. Zep and Cognee funding amounts are undisclosed; treat their runway as unverified. Supermemory's Google/Nissan deployments and benchmark leads are vendor-claimed only. Letta's 24k-star repo is now labeled the legacy V1 server — star count overstates current-product momentum. Adjacent elements: vector databases (pgvector, Pinecone, Turbopuffer) → Vd; company wikis and 'company brain' knowledge bases (incl. Hyperspell's overlap) → Kw; meeting memory (Granola, Otter) → Mt; MCP servers as distribution → Mc. Native memory in Claude/ChatGPT is covered here only as the bundling threat, not as picks — it doesn't cross apps or models, which is this element's whole job. Ranking criteria: verified commercial traction, independent satisfaction surveys, agent benchmarks, and founder-fit (price floor, lock-in, surfaces). Editorial, never paid — the charter. Machine-readable twin: mm.json.

Our take

Nascent and worth watching. The first element where 'my agent knows my business' becomes literal.

Combines with

Appears in compounds

The Agent Stack