Daniel Chalef

@krshgl @zep_ai Sorry about that!

The funniest part is that this isn't even something I kept saved up for the launch, immaculate timing. Also a kind request for @zep_ai to please index the citations correctly, liddle liddle boosts like these are so fun. arxiv.org/abs/2501.13956

I spent my first month @LittlebirdAI dissecting existing memory systems - @zep_ai's paper among them. Today, during a random testing session around personal understanding, LB told me that my first paper was cited in the Zep paper. I track my citations 😁 (because there are so few) and this wasn't indexed so I was 100% sure that this is a context hallucination. Dug deeper and found out that @mahek_chandak and I are actually cited in the Zep Temporal KGs paper. So crazy.

@ShubhamNahar @swyx @emileifrem @lyonwj Zep does this today.

@swyx @emileifrem @lyonwj the reasoning memory piece is what separates this from Mem0/Zep. actually recording the why behind decisions, not just the what

just learned about create-context-graph - one command and it sets up key entity relationships for 22 top industry domains. been looking for something like this to "layer on" a social graph to every single app I make!

Every agent context architecture comes down to three decisions: → Scope: per-user context, shared domain knowledge, or both? → Data: conversations, business data, or both? → Retrieval: deterministic assembly or agent-controlled? Get these right and everything else follows 🧵

Had a ton of "Claws" sign up for @zep_ai overnight 🦞. Looks like MiroFish uses Zep for persistent context. Cool project! github.com/666ghj/MiroFish

Most teams shipping AI agents can't verify context retrieval actually works for their domain. We built a framework to answer that. Also: a new Zep CLI and dashboard overhaul. 🧵

We've been heads down shipping updates to Zep's context graph platform. Here's part 1 of our updates: → Custom extraction instructions → Property & exclusion filters for graph search → Webhooks What's new in building and querying context graphs 🧵

@meta_alchemist . @meta_alchemist Graphiti has 22K+ stars, way more than Cognee and Letta... Would you please correct your chart? Also, @zep_ai has raised way more funds 😉

AI normally sucks at memory retrieval both you and your AI hate that Here are the top 10 open source AI memory layers > free > open source > with lots of GitHub stars > and some even funded by YC You can use them to make your agents, claude, codex, and openclaw much smarter Also, tips on what each memory is good at, how to combine them, and get even more out of them are all in the article. These are much better than just memory md files, and will make a huge difference in your workflow! Give your agents life with them:

Code reviews at @zep_ai 😂

My wife maintains a 35,000-token "Master Prompt" in Notion. 🤯 She manually copy-pastes her entire medical history, life events, and goals into every new AI chat just to get personalized advice. So I built an AI Chat with "Deep Memory" powered by @convex and Graphiti by @zep_ai

@jezell @ladybugmem @zep_ai Will do. Thanks

@ladybugmem Yo @zep_ai update your docs. Kuzu is dead.

Ladybug Memory is now composed of multiple public projects. Please join the discord and subscribe to project specific channels. Invite on the website as well as in the replies.

@DenissAlimovs @zep_ai Sorry we missed this. Webhooks are available with Zep's FlexPlus plan: getzep.com/pricing/

@zep_ai I just purchased Flex plan but can't see Webhooks in my project on left hand side as described in docs. What am I missing?

@saleban1031 Those Zep numbers are incorrect. Zep is 80.06% at 241ms retrieval latency with 1,378 context tokens. It's not just raw accuracy that counts, but also latency and token expense. blog.getzep.com/the-retrieval-…

After extensive testing, Memvid hit 85.7% on LoCoMo, the hardest long-term memory benchmark. That’s a 28% relative gain over Mem0, OpenAI Memory, Zep, and other systems. - No vector database. - No RAG pipeline. - No infrastructure to maintain. - No config sprawl. Just conversations preserved as-is, encoded into a single memory file. Most AI memory systems fail because they throw away structure. Open source + fully reproducible: github.com/memvid/memvidb… Full report: memvid.com/benchmarks

@logonx2421 @deepacodex What’s a “boomer product”?! 🤔

@deepacodex this weekend: - wiring up memory to claude code - tried supermemory, onboarding is broken - tired zep, seems like boomer product for b2b customers - settled on mem0 - claude code hooks are so cool, you should learn them - on session start claude will know everything about me

Its Friday 🎉🎉 What are you building this week? Share links, curious to see.

@hafezverde @pontusab @nicoalbanese10 @aisdk @zep_ai And for those new to Zep: getzep.com

@hafezverde @pontusab @nicoalbanese10 @aisdk You can do this with @zep_ai 's Standalone Graphs: help.getzep.com/create-graph

Excited to announce that AI SDK Tools now support @aisdk v6! - `ai-sdk-tools/agents` - `ai-sdk-tools/artifacts` - `ai-sdk-tools/cache` - `ai-sdk-tools/memory` - `ai-sdk-tools/store` It's under the v2 beta tag.

@dilipsrajan @DhravyaShah @supermemory Founder of Zep here. What were your concerns regarding Zep’s graph implementation? How does it impact the DX for you as a developer?

@DhravyaShah @supermemory spent a lot of time this summer experimenting with memory architectures and came to the same conclusion mem0g and zep have good ideas but too bogged down in trying to be a graph it's way more practical to be graph-ish

The graph architecture behind @supermemory came from pure spite. We freaking hated graphs. Our solution? Don't build one. Thought from first principles instead: - Memories update - New ideas extend old ones - Everything relates (1-1, 1-many, many-1) Turns out that's a graph lmao BUT our version is different. No triplets. No entity-relation-entity parsing. Just facts. "dhravya likes pizza" stays whole. Not broken into (Dhravya, likes, pizza) nodes. On query time, the result directly hits, instead of traversing through thousands of nodes to build up the context When preferences change, it links to "Dhravya likes pasta" via an update relation. Dead simple. Facts on facts. That's the whole thing.

@0xTechDean You may want to check out Graohiti’s MCP server for this. Knowledge Graphs for agent context: github.com/getzep/graphiti

Built a mem0 plugin for Claude Code. Thought why not Cursor too? Persistent memory for your AI coding assistant: * Retrieves relevant context before each prompt * Saves conversations between sessions * Open source, simple hooks setup PRs welcome github.com/0xtechdean/cur… @cursor_ai @mntruell @amanrsanger @mem0ai @taraborself

@code_kartik @dishantwt_ And @zep_ai 😉

@dishantwt_ Supermemory and mem0 are doing this only ig.

how about creating a context tree for long conversations so that models only get the particular set of information by navigating to a particular node trees can form multi-level topic hierarchy, so long raw conversations can be easily represented without summarising and having lossy compressions

#MCP for agent memory: @zep_ai Graphiti + FalkorDB for persistent, multi-tenant knowledge graphs. youtu.be/tjEznowIhBo

This is such a ridiculous take on a well-known problem. If the vector DB is breached, the source content is likely breached too. There’s nothing novel here. Secure the whole service properly.