llms.txt
Compare with other sites' llms.txt files →# Papr
> Papr builds context intelligence for AI agents (Papr Graph: memory with vector embeddings + knowledge graphs) and Papr Work, a local-first desktop app where AI builds and runs small apps and automations for go-to-market teams.
## Product
- [Papr Work](https://www.papr.ai/landing): Desktop app (Mac, Windows) for AI apps that run GTM work: prospect research, CRM updates, outbound, meeting prep, reporting.
- [Papr Graph](https://www.papr.ai/landing/graph): Memory and context API for AI agents; Python and TypeScript SDKs and MCP.
- [How it works](https://www.papr.ai/landing/how-it-works): How Papr memory and context retrieval work.
- [Pricing](https://www.papr.ai/landing/pricing): Plans and limits.
## Blog
- [Papr Blog](https://www.papr.ai/blog): Launches, guides and research notes.
- [Small software is a team sport now](https://www.papr.ai/blog/small-software-team-sport): Papr Work 2.7.3 adds Discover, Connections for the tools you use, an Auto model with token tracking per app, and roles for every shared app.
- [Why every teammate gets their own copy of a shared app](https://www.papr.ai/blog/work-on-apps-together): Papr Work gives every teammate their own linked copy of a shared app instead of edit rights to one file. Teammates send changes back to the owner as proposals, the owner decides whose API keys the app uses, and Get updates brings in the owner's changes without losing your own edits.
- [Introducing papr-embed: schema-native embeddings, and open-weights embeddings for code retrieval](https://www.papr.ai/blog/introducing-papr-embed-schema-native): Papr released two embedding model series. papr-embed-code-1 (4B and 0.6B, open weights under Apache-2.0) ranks #1 on all four MTEB code leaderboards, Code, CoIR, RTEB Code and CodeRAG, mostly zero-shot, and beats F2LLM-v2-14B on CoIR at under a third of its size. papr-embed-schema-1 conditions embeddings on your data schema, and its 4B beats Qwen3-Embedding-8B on SciFact.
- [Stop Being the Glue Holding Your GTM Stack Together](https://www.papr.ai/blog/stop-being-the-glue-holding-your): Most teams have a CRM, Slack, email, docs and AI chat, yet a person still moves the work between them. In this free live webinar on Thursday, October 29, attendees bring one task they never want to do by hand again and we build a working workflow for it live in Papr Work.
- [I Built a Papr Work App to Manage Childcare for Seattle Kraken Game Nights. It Took One Conversation.](https://www.papr.ai/blog/i-built-an-app-to-manage-childcare): Kraken Sitter is a personal Papr Work app that lets babysitters claim Seattle Kraken home game nights from a preloaded season schedule, with confirmations sent from the parent's own Gmail. It was built by describing it to the agent in four sentences, and later features like a kid count and sitter notes were added the same way. It shows how easy it is to build small software for everyday problems.
- [LinkedIn Outreach: Spend Your Time on Clients, Not a New Side Project](https://www.papr.ai/blog/linkedin-outreach-spend-your-time): LinkedIn Outreach is a ready-made Papr Work app that runs multi-step LinkedIn campaigns through your own account, with prospects, sequences, follow-ups and a reply inbox working from day one. It is for solo consultants and small GTM teams who want more client conversations without designing and maintaining an outreach automation themselves.
- [The Craft Behind Every Design Decision at Papr](https://www.papr.ai/blog/the-craft-behind-every-design-decision): Chat should not be the main interface for AI: the work should happen in the background and arrive as small software you can inspect and own. We rebuilt Papr Work's navigation around three questions, where am I, is my agent working, and does anything need me, cut the sidebar from 240px to 96px, and took the logo through five versions to a single infinity stroke that also forms the agent's eyes.
- [Papr Work, Redesigned Around the Way You Work](https://www.papr.ai/blog/papr-work-redesigned-around-the-way): The redesigned Papr Work has a slimmer interface that gives the work more room, Focus to keep three priorities in view, Replay to show what your agent is doing while it works, and sharing so a useful app can travel to teammates. It is built for people who use AI but still coordinate work across many tools.
- [GTM Unblocked Is Nearly Full: We Added 12 More Spots.](https://www.papr.ai/blog/gtm-unblocked-is-nearly-full-we-added): We opened 12 more spots for GTM Unblocked, our hands-on GTM automation workshop on September 9 at the AWS Builders Loft in San Francisco. Bring one workflow you are tired of doing by hand, such as meeting prep, prospect research or CRM updates, and leave with it running in Papr Work.
- [GTM Unblocked: A Hands-On AI Automation Workshop](https://www.papr.ai/blog/gtm-unblocked-a-hands-on-ai-automation): GTM Unblocked is a hands-on working session on September 9 at the AWS Builders Loft in San Francisco for founders, marketers, account managers and agency operators. Attendees bring one repetitive GTM task, such as CRM updates, meeting prep or follow-ups, and use Papr Work to turn it into an AI app that runs it for them.
- [Papr Launches Papr Work to Replace Fragmented GTM Software With AI Apps That Run the Work](https://www.papr.ai/blog/papr-launches-papr-work-to-replace): Papr Work is a local-first desktop app for founders, marketers, account managers and independent professionals at small and midsize businesses. It replaces repetitive go-to-market work across Gmail, Slack, Calendar, CRMs and LinkedIn with AI apps you can run as-is, customize or build in chat, and the apps keep persistent business context so they do not start from scratch.
- [Stop Building Fake Employees. Automate the Work You Hate.](https://www.papr.ai/blog/how-to-build-great-agents-stop-trying): AI agents usually fail because they are built as fake employees meant to replace a whole function. The agents that work start narrow: they automate one task someone hates or replace one narrow SaaS tool, fail loudly when they are wrong, and grow from there. The post walks through that playbook with community-built examples from Papr Work.
- [Stop Engineering Context. Engineer the Loop.](https://www.papr.ai/blog/loop-engineering-the-next-frontier): Loop engineering means designing the outer loop an agent runs in, not just its prompt or context, because loops are what produce improvement. We built one for our own ML research: a controller that proposes one experiment per iteration, runs it in an isolated git worktree, scores it on a fixed objective and keeps only what wins. The post shares five patterns from it.
- [Why RAG Fails Before the Model Gets Involved](https://www.papr.ai/blog/why-rag-fails-before-the-model-gets): Many production RAG failures happen at retrieval, not generation: the agent gets the wrong context before the model starts writing. Adding more context makes this worse as data grows, so the retrieval layer needs structure, such as entities and relationships, to return the right context instead of the most similar text.
- [The Trillion-Dollar Architecture Shift: Why AI Agents Need Context Graphs, Not Just Vectors](https://www.papr.ai/blog/why-vector-search-alone-will-never): Vector embeddings find similar text, but agents need to know which entities are the same, how they connect and what is allowed. Papr's memory policies and graph constraints, shipped in the Python SDK v2.21.0 and the TypeScript SDK, turn memory into a context graph with one canonical entity per concept, queryable through natural language or GraphQL, with safety enforced on the memory itself.
- [Friday Night Experiment: I Let a Multi-Agent System Decide Our Open-Source Fate. The Result Surprised Me.](https://www.papr.ai/blog/i-built-a-multi-agent-ai-to-decide): To decide whether to open-source Papr's predictive memory layer, we built a multi-agent reinforcement learning system with four stakeholder agents backed by Papr memory, then ran 100,000 Monte Carlo simulations and 10,000 training episodes. 91.5% of runs favored open-core, with an average NPV of $109M versus $10M for proprietary, and agents with deeper memory favored open-core more strongly. We open-sourced the core.
- [Beyond RAG: Introducing Papr Context Intelligence](https://www.papr.ai/blog/beyond-rag-introducing-papr-context): Papr Context Intelligence builds on Papr's predictive memory so agents can reason over what they remember, not just look it up. It connects memories by cause and impact, produces insights and analytics through natural language or GraphQL, and learns what matters from real usage, replacing the usual stitched-together vector database, knowledge graph and chat history.
- [Introducing Papr Predictive Memory](https://www.papr.ai/blog/introducing-papr-predictive-memory): AI agents forget everything between sessions, and adding more data to a vector store often makes retrieval worse. Papr Predictive Memory measures this with a retrieval loss metric and predicts the context an agent will need before it asks, combining vector embeddings and knowledge graphs. Developers can use it through Python and TypeScript SDKs.
- [Memory Networks: Building AI That Remembers Like Your Brain](https://www.papr.ai/blog/memory-networks-building-ai-that): Knowledge workers lose about 2.5 hours a day searching for information spread across tools, and AI assistants make it worse by forgetting context between conversations. Papr's memory network gives AI a brain-like memory across your workspace and took the #1 spot on Stanford's STARK retrieval benchmark.
## Optional
- [RSS feed](https://www.papr.ai/blog/rss.xml)
- [Sitemap](https://www.papr.ai/sitemap.xml)