Cognition Buys Poke for Memory, Not Personality

Cognition bought messaging agent Poke for a reported low-nine-figure valuation, at least four times the $25 million its maker had raised. The asset is not a charming chatbot: it is persistent memory and proactive coordination proven across more than 100 million messages in three months, capabilities Cognition wants to graft onto Devin.

The official Cognition announcement says Poke will remain available without immediate changes through the end of 2026 while experiments move into Cognition’s products in 2027. TechCrunch confirmed the low-nine-figure valuation, though neither company disclosed a purchase price or cash-stock split. That leaves operators with a strategic signal, not a shipping feature.

The premium is for continuity

Interaction Company had only 10 employees, had raised $25 million, and carried a $300 million post-money valuation before the transaction, according to TechCrunch’s April product profile. “Low nine figures” begins around $100 million. Divide that floor by disclosed capital and Cognition paid at least 4× funding raised. Divide the same floor by Poke’s latest 100-million-message quarter and the deal assigns roughly $1 of acquisition value per recent message. Neither ratio is a revenue multiple—Poke acknowledged difficult unit economics—but both show Cognition buying interaction data and habits, not headcount.

Poke works through iMessage, SMS, Telegram, and WhatsApp. It messages first, follows up, remembers previous work, chooses models by task, and connects to tools through recipes and MCP. Its first-party MCP documentation lets other software send messages into that relationship. The behavior sounds consumer-oriented until applied to software delivery: a persistent layer can remember why an earlier Devin session made a choice, coordinate several sessions, and return to the engineer when a dependency clears.

That is where current agents remain clumsy. Generating a patch is becoming cheap; assigning ten asynchronous tasks, reconciling their assumptions, noticing one stale branch, and restoring context next week is not. Cognition already supports parallel sessions, Slack and Teams initiation, schedules, MCP, and Auto-Fix. Its own guidance on when to use Devin favors well-scoped, verifiable tasks. Poke’s value is whether it can keep those tasks coherent after the scope crosses a session boundary.

Cognition supplies the execution engine Poke lacked. The company says Poke can move onto SWE-1.7, its model optimized for asynchronous long-horizon work. The SWE-1.7 announcement describes self-compaction and training trajectories up to six hours. Pair that with a messaging layer that maintains longitudinal state and the product thesis becomes legible: Poke owns the ongoing relationship; Devin executes the engineering work.

This is a different approach from simply making a computer-use model cheaper, the bet behind Prentis’s low-cost agent lab. It also complements the long-context economics in Kimi K3’s million-token routing proposition. Context windows hold information inside a run. Persistent product memory decides what should survive between runs and when to act on it.

Test coordination, not charm

Enterprises should not switch coding agents because Poke feels personable. The integration is not generally available, no benchmark measures it, and persistent memory can preserve stale assumptions or sensitive context as easily as useful history. A colleague-like voice can also hide uncertainty: an agent that sounds familiar may receive less scrutiny precisely when it needs more.

Unit economics are another fault line. Poke lists free, $19-per-month Pro, and $199-per-month Ultra plans, while public Devin pricing starts at $20 per month for Pro and $200 for Max, with an $80 Teams minimum and negotiated enterprise usage. There is no announced acquisition surcharge. The real migration bill includes repository indexing, permission design, memory retention policy, security review, and human PR review. Poke’s earlier admission that proactive service was expensive means inference discipline must improve under Cognition rather than merely move balance sheets.

The right buyer is a team already losing engineer time to agent coordination: repeatedly translating Slack threads into tasks, restoring repository context, monitoring concurrent runs, and chasing follow-ups. Run a four-to-six-week matched pilot once the integration ships. Compare cost per merged PR, human coordination minutes, re-prompt frequency, cycle time, rollback rate, and security exceptions against the current workflow. If “personality” cannot move those numbers, it is decoration.

The thesis breaks if the integration remains a 2027 experiment, if memory crosses repository or customer boundaries, or if added proactivity increases actions without increasing accepted work. Messaging-platform dependence is another risk; Apple approval reportedly took months, and platform rules can change faster than enterprise roadmaps. A production incident involving stale or over-broad memory would reverse the trust benefit instantly.

Evidence that changes the verdict is straightforward: generally available persistent orchestration, lower human coordination time, more merged work per engineer, stable defect rates, and a transparent memory-control plane. Until those arrive, teams should keep evaluating Devin on execution rather than acquisition narrative. Cognition’s purchase says the next agent bottleneck is continuity. It has not yet proved that Poke solves it.

The 4× minimum valuation-to-funding ratio is the market’s bet on that solution. It is also a useful warning: when ten people and 100 million messages command nine figures, the scarce asset in agentic engineering is no longer another code generator. It is the product layer that remembers what all the generators were supposed to accomplish.