Cover Analysis · N°01

The State of AI Memory — Q2 2026

Four categories, one missing layer, and why MCP just changed the landscape.

The phrase “AI memory” is doing too much work. In 2026 it covers four distinct markets with different buyers, different price points, and different competitive logic. Treating them as one category is the fastest way to misread the landscape — for builders, for investors, and for anyone trying to figure out where their data actually lives.

This piece maps the four categories, the money flowing into each, and the gap that we believe is opening up between them.

Four categories, not one

Category	Representative players	What’s sold	Buyer
Agent memory infrastructure	Mem0, Letta, Zep, Cognee, Cloudflare Agent Memory	APIs and SDKs that give other developers’ agents persistent state	Engineers building agents
Walled-garden memory	ChatGPT Memory, Anthropic memory features	Per-vendor recall inside one chat product	End users of one product
Consumer “second brain”	Mem.ai (original), Notion AI, several smaller note-AI hybrids	Notes plus AI on top of them	Knowledge workers
Personal AI companions	Character.AI, Replika, Inflection’s Pi (now Microsoft)	A continuous relationship with a single AI persona	Consumers seeking a companion

Each category is real, well-funded, and answering a different question.

Agent memory infrastructure: B2D, well-capitalized, growing fast

This is the most active category in 2026.

Mem0 raised a $24M Series A in late 2025 led by Peak XV and Basis Set, with YC participation. Their open-source repo crossed 41,000 GitHub stars and 13M PyPI downloads — extraordinary numbers for a memory-layer library, and a leading indicator of how many developers are now treating memory as a separate concern from the model itself.

Letta (emerged from stealth in Sept 2024 with a $10M seed at a $70M post-money valuation, led by Felicis with angels including Jeff Dean and Clem Delangue) descends from the MemGPT work at Berkeley’s Sky Computing Lab. Their 2026 push has been around Context Repositories — a programmatic, git-versioned model of agent memory — and an open-source coding agent that ranks #1 model-agnostic on Terminal-Bench.

These are developer tools. The buyer is an engineering team building an agent. The price point is per-API-call or per-seat for a platform. Neither company is trying to ship an end-user product, and that distinction matters when you read about “the AI memory market” — most of the activity is here, but most of the activity is not about you and your AI.

Walled-garden memory: dominant by default, captive by design

ChatGPT’s memory feature is now standard across paid tiers. It works in two modes: saved memories you explicitly request, and chat history — implicit context from past conversations. Per OpenAI’s Memory FAQ, it can also pull from connected apps like Gmail and a Plus user’s file library.

Anthropic and others have shipped equivalents. The pattern is consistent: useful, well-integrated, and architecturally tied to the product. Memory in ChatGPT lives inside ChatGPT. Memory in Claude lives inside Claude. There’s no expectation — and no plumbing — for it to travel.

For most people this is invisible, because most people use one AI. But the population that uses two or more is large and growing, and for that population the walled-garden model produces a specific failure: re-explaining. Every new app, every new model, every new chat begins from zero. The cost is small per instance and large per week.

Consumer “second brain”: the cautionary category

Mem.ai (the original 2019 note-taking startup, not to be confused with Mem0) raised over $29M from Andreessen Horowitz, Kortschak, Neo, and OpenAI Startup Fund. In 2025 a widely-circulated post-mortem framed the company as a “$40M Second Brain failure.” The premise — AI sitting on top of your personal notes — turned out to be harder to make sticky than the deck suggested.

The pattern is broader than one company. Several “AI plus your notes” products have struggled to find retention even when they find adoption. Workflows are sticky; new workflows are not. The lesson the market took: don’t ask the user to start a new habit; meet them where their AI already is.

Personal AI companions: enormous market, winner-take-all

The companion market is the loudest one. Character.AI — 233 million registered users — was acquired by Google for $2.7B. Replika reports over 30M users. Inflection’s Pi went to Microsoft for $650M. MIT Technology Review named AI companions one of its 10 Breakthrough Technologies of 2026.

Sizing forecasts vary widely depending on definition: Precedence Research projects $49B (2026) → $552B (2035). The mobile-app slice generated $82M in H1 2026, on track to exceed $120M by year-end with 88% YoY download growth. But the distribution is brutal: top 10% of apps capture 89% of revenue. ~33 apps, plus 300+ struggling.

This category is real but is not, despite frequent confusion, the same market as memory infrastructure. Replika sells emotional continuity. Mem0 sells a Python package. Conflating them produces bad strategy in both directions.

What changed in 2026: MCP became infrastructure

The single biggest shift this year was Model Context Protocol crossing into “infrastructure” territory.

MCP was introduced by Anthropic in November 2024 as an open standard for exposing structured context — files, tool results, state — to LLMs through a uniform JSON interface. In December 2025 Anthropic donated it to the Agentic AI Foundation, a Linux Foundation directed fund co-founded with Block and OpenAI. By Q1 2026 it had reached 97 million monthly SDK downloads and become the de facto plumbing for AI-to-tool connectivity.

Two consequences are already visible:

Memory APIs are rapidly becoming MCP servers. Mem0, Letta, and others are exposing themselves through MCP so any compliant client can connect. The integration tax that used to live at the application layer is dissolving.
The barrier to cross-vendor memory has collapsed. Until 2025, “memory that works across ChatGPT and Claude and Cursor” was a heavy engineering project with bespoke integrations per client. With MCP it is — at the protocol level — a configuration question. Every major AI client now speaks MCP or is in the process of adding it.

What MCP does not solve: who owns the memory, where it’s stored, and whether the user controls it. Those are product and policy questions. The protocol just made them answerable.

A Möbius loop in violet against black — continuous, with no beginning and no end.

The gap

Holding the four categories and the MCP shift in view, a specific gap shows up.

There is no widely-adopted consumer product whose value proposition is “the memory that travels with you across every AI you use.” The infra layer is there. The standard is there. The market need — measured by the volume of “I had to re-explain X to ChatGPT again” complaints in any developer forum — is there. What hasn’t been built is the user-facing product that owns the portability claim.

The companion apps don’t fill it: their model is single-vendor by design, because the persona is the product. The walled gardens don’t fill it: portability is structurally against their interest. The agent-memory startups don’t fill it because they aren’t end-user products. Mem.ai didn’t fill it because notes-plus-AI is a different shape entirely.

This isn’t a prediction about who should fill the gap. It’s an observation that, for the first time, the gap is structurally fillable — the protocol gravity is there, and consumer awareness of “AI amnesia” is mainstream enough that the pitch lands without education.

We expect 2026–2027 to settle whether someone takes that ground or whether the walled gardens dissolve the demand by getting good enough at intra-vendor memory that users stop noticing the seams.

Disclosure: iCog is building in this space. We’ve tried to keep this piece descriptive rather than positioning. Where you read editorial sharpness, assume bias.