Three New AI Memory Tools You've Never Heard Of, Honestly Compared

Search for the best AI memory tools and the same shortlist comes back every time. Underneath it, a newer wave is quietly shipping, and three of them are worth a closer look because each bets on a different answer to the same question: what should an AI remember, and where should that memory live?

Unabyss bets on your work apps. Spellar bets on your meetings. CogniMemo bets on an API your own products call. And there is a fourth bet, the hosted memory layer, where one memory follows every tool and scheduled agent you run. This comparison walks through all four: what each remembers, what each costs, and which job each is built for.

Start with the job, not the ranking

Memory sounds like one feature. In practice it is at least four different problems wearing the same name. Remembering what your apps contain is one problem. Remembering what was decided on calls is another. Giving an application you are building a memory store is a third. Making sure a scheduled agent that runs at 2 a.m. knows what its 9 a.m. run learned is a fourth. A tool can be excellent at one of these and not designed for the others, which is why ranking pages that crown a single winner mislead. The useful question is never "which tool is best?" It is "which job do you need done?"

Unabyss: memory built from your work apps

Unabyss builds a context vault out of the apps where your work already lives. Connect sources such as Gmail, Slack, Notion, GitHub, Google Drive, and your calendar, and it extracts and structures what they contain into a curated picture of your projects, preferences, and working style. That vault is then served over MCP, so compatible tools like Claude, Cursor, and Codex can pull from the same context instead of starting cold.

Two design choices stand out. First, agents read from the vault rather than reaching into your live accounts, and access is controlled through four permission scopes with named, revocable tokens. Second, pricing is usage based: you start with $5 in free credits and pay as you go, with no tiers to decode. Setup runs through OAuth, roughly two minutes per integration.

Unabyss fits knowledge work spread across many apps, where the pain is re-explaining your world to every new tool. It shines when the answer already exists somewhere in your documents, messages, and files.

Spellar: memory built from your meetings

Spellar takes the opposite slice of your working life: the calls. It is a meeting companion for macOS, iPhone, iPad, and the web that records without a bot joining the call, transcribes on device, and builds memory across meetings, so decisions, commitments, and action items stay connected from one call to the next instead of scattering across separate note files.

You can ask questions across your meeting history through its query interface, and push summaries and action items into tools like Notion, Linear, Jira, Slack, and Google Docs. Server-side AI processing is opt-in, you can bring your own API key, and the product is built under UK GDPR and CCPA. Pro costs $18.99 per month on monthly billing, or $11.99 per month when billed annually.

Spellar fits anyone whose real work happens in conversations: sales calls, client check-ins, team standups. If the context you keep losing is "what did we agree on last time we spoke," a meeting-native memory is the direct answer.

CogniMemo: memory as an API

CogniMemo is aimed at builders rather than end users. It offers memory infrastructure through a single API and SDK: store user preferences, tasks, decisions, and conversation patterns once, and any model your product uses, whether that is OpenAI, Claude, Gemini, or a local model through Ollama, can read and write the same memory. It connects to storage backends including Pinecone, Weaviate, PostgreSQL, and Redis, and offers LangChain adapters, with encrypted, permission-based access that users can revoke.

The product is currently in free beta, with production pricing still to be announced. If you are building an app whose users expect it to remember them across sessions, you can prototype against a real memory API today and decide about production once pricing lands.

The fourth bet: one hosted memory for every tool and agent

The hosted memory layer takes a different position from all three. Instead of extracting memory from one source, or exposing an API for one product, it sits outside every tool as a single shared store. Every connected client reads and writes the same memory over MCP, so a decision made in a desktop chat in the morning is available to a phone app at lunch and to a scheduled automation overnight.

Vilix AI is built on exactly that model. It is cloud-hosted, so there is zero infrastructure to install, run, or maintain. The same memory is available everywhere over MCP, on every compatible tool and device you connect. It stores full conversation history, not just extracted facts, so the actual conversation can be revisited later, including the reasoning behind a decision, not only the decision itself. For automation operators, that means a recurring agent can check what previous runs did, what they found, and what changed, instead of waking up blind and re-deriving everything from scratch.

The entry terms are deliberately low friction. The free plan is free forever. A 7-day Pro trial is available with no credit card required. And your data stays yours: you can export everything or delete it at any time, whenever you want.

Side by side

Picking in under a minute

If your problem is that every new AI tool knows nothing about your work, and that work lives in your apps, start with the context vault model. If your problem is losing what was said and decided across calls, start with the meeting companion. If you are shipping a product and need memory inside it, start with the API while its beta is free. And if your problem is that your tools and scheduled agents each start from zero while you re-brief them every day, start with one hosted memory they all share.

The best AI memory tool is the one matched to the job you actually have. Name the job first, and the shortlist sorts itself.