Releases and product changes, newest first. Where a change touches retrieval, the entry says what it measured — a memory product that claims recall got better and shows nothing is asking you to take its word for it.
Chain recall got substantially better on the default settings
Cogni’s recall seeds a graph traversal from the memories that match your question directly, then follows the connections outward. Those starting points were allowed to fill the whole response, so on the questions the graph exists for — where the answer lives in a memory that shares no words with the question — the traversal’s results often never reached you.
Seeds are now bounded to half the response. Measured through the real server on the default mode and default result count, with a frontier model answering: 0.06 → 0.48 on cross-vocabulary multi-hop questions, and 0.00 → 0.45 on the two-hop subset. It also costs 26% fewer internal searches.
Shipped alongside it: retrieval breadth now scales with how much you have stored, so recall quality holds as a store grows rather than quietly thinning out at the top of the plan.
A team space is a shared memory with its own store and its own membership — separate from anyone’s personal memory, not a folder inside it. Creating one, adding and removing people, and seeing who has access now all happen from your account page instead of requiring us.
Two rules worth knowing: administration is limited to a space’s owners, and a space can never be left with no owner — the last one cannot be removed.
Creating a team space needs a Team plan; managing an existing one does not.
Launch
Cogni is live
Sign-up is open. Cogni is an MCP memory server: connect it to Claude, ChatGPT, Cursor, Claude Code or your own agent, and it gives that model a memory that persists between sessions and follows the connections between what you have told it.
The part people are usually surprised by is that it is one memory. Sign in on Claude Desktop, mint a key for your terminal, add the connector in ChatGPT — all three read and write the same store. Your agent remembers the same things everywhere you work.
Free needs no card. There is no vector database to run, no GPU, and no LLM anywhere in the retrieval path — nothing rewrites or summarises your memories on the way in, and nothing is generated on the way out.