Hey all, I'm one of the founders at Papr.

We started building an AI project manager. Users needed to search for context about projects, and discover insights like open tasks holding up a launch.

Vector search was terrible at #1 (couldn't connect code, marketing and PR that are for the same project). Knowledge graphs were too slow for #1, but perfect for structured relationships, great for UIs.

Then we started talking to other teams building AI agents - we realized everyone was hitting the exact same two problems.

So we pivoted to build Papr — a unified memory layer that combines: - Intent vectors: Fast goal-oriented search for conversational AI - Knowledge graph: Structured insights for analytics and dashboard generation - One API: Add unstructured content once, query for search or discover insights

And just open sourced it.

How intent vectors work (search problem) The problem with vector search: it's fast but context-blind. Returns semantically similar content but misses goal-oriented connections.

These are far apart in vector space (different keywords, different topics). Traditional vector search returns fragments. You miss the complete picture.

Our solution: Group memories by user intent and goals stored as a new vector embedding (also known as associative memory - per Google's latest research).

When you add a memory: 1. Detect the user's goal (using LLM + context) 2. Find top 3 related memories serving that goal 3. Combine all 4 → generate NEW embedding 4. Store at different position in vector space (near "product launch" goals, not individual topics) 5. Query "What's the status of mobile launch?" finds the goal-group instantly (one query, sub-100ms), returns all four memories—even though they're semantically far apart.

This is what got us #1 on Stanford's STaRK benchmark (91%+ retrieval accuracy). The benchmark tests multi-hop reasoning—queries needing information from multiple semantically-different sources. Pure vector search scores ~60%, Papr scores 91%+.

Automatic knowledge graphs (structured insights) Intent graph solves search. But production AI agents also need structured insights for dashboards and analytics. The problem with knowledge graphs: - Hard to get unstructured data IN (entity extraction, relationship mapping) - Hard to query with natural language (slow multi-hop traversal) - Fast for static UIs (predefined queries), slow for dynamic assistants

Our solution: - Automatically extract entities and relationships from unstructured content - Cache common graph patterns and match them to queries (speeds up retrieval) - Expose GraphQL API so LLMs can directly query structured data - Support both predefined queries (fast, for static UIs) and natural language (for dynamic assistants)

We combined both of these solutions in one API.

What I'd Love Feedback On

1. Evaluation - We chose Stanford STARK's benchmark because it required multi-hop search but it only captures search, not insights we generate. Are there better evals we should be looking at?

2. Graph pattern caching - We cache unique and common graph patterns stored in the knowledge graph (i.e. node -> edge -> node), then match queries to them. What patterns should we prioritize caching? How do you decide which patterns are worth the storage/compute trade-off?

3. Embedding weights - When combining 4 memories into one group embedding, how should we weight them? Equal weights? Weight the newest memory higher? Let the model learn optimal weights?

4. GraphQL vs Natural Language - Should LLMs always use GraphQL for structured queries (faster, more precise), or keep natural language as an option (easier for prototyping)? What are the trade-offs you've seen?

---

Try it: - Developer dashboard: platform.papr.ai (free tier) - Open source: https://github.com/Papr-ai/memory-opensource - SDK: npm install papr/memory or pip install papr_memory

Using .gov Email Addresses for Age and Information Verification

Report ITU-R TF.2511-0 (10/2022): Content and structure of time signals to be di [pdf]

AmneziaWG – WireGuard-Go Fork, Avoid WireGuard Deep Packet Inspection Detection

Show HN: Oh-My-Opencode: Async Subagents, Curated Agents. $24,000 Worth Spent.

Sloot Digital Coding System

What was typography like in the Soviet Union? (A Bluesky thread.)

The "Homeless" Streamers of China [video]

Prediction Markets Are Accurate

What are the chances the cashless student cafeteria has a privacy policy?

Why Young People Are Struggling to Communicate

Today Is Zamenhof Day

Sound Waves Can Fight Fires Without Water

Reliza and ShiftLeftCyber Announce Integration of SecureSBOM Signing into ReARM

Show HN: Intent vectors for AI search and knowledge graphs for AI analytics

A rather simple text editing tool. Try it out

Show HN: Octopii – runtime for writing distributed applications in Rust

My Thoughts on AI Safety

Should anti-bullying approaches encourage kids to be 'upstanders?'

Show HN: A practical, code-driven guide to learning HTTP frameworks in Go

Drawing parallels between home renovation and software development

Marvell Has a 2nm Ultra-Power-Optimized Memory – ServeTheHome

Hacking Endpoint to Identity: "ConsentFix" [video]

Rust Is Officially Part of Linux Mainline

Clean, Limitless Energy Exists. China Is Going Big in the Race to Harness It

Timing 'Hello, World'

Nvidia Debuts Nemotron 3 Family of Open Models

Virtualizing Nvidia HGX B200 GPUs with Open Source

Adeptly

Hype Correction

Picture book maker using AI Agent

Show HN: I built a fast cross-platform CLI network scanner via Python

Comments