.@zep_ai introduced Zep Agent Skills, and it’s really worth your attention.

It's a procedural memory system that turns verified, successful AI agent runs into reusable Skills.

The idea of procedural memory isn't new, but Zep is trying to turn it into a managed service, with a ting details ↓

First, here's how it works:

1. Tracing: Zep records the agent's actions, tool calls, and outcomes as a Trace. It can also import Traces from Braintrust and LangSmith.

2. Verification: Checks whether the task actually succeeded using independent evidence, not just the agent's claims.

3. Compilation: A Reflector proposes reusable instructions, while a Curator checks them against the evidence. Each Skill field links back to specific Trace events.

4. Admission: Safety is a very important point. Security scans and quality gates determine whether a new Skill version becomes available.

5. Retrieval: Agents find relevant Skills through lexical search, load instructions when needed, and record outcomes that influence future rankings.

Then, there are a few features that are particularly interesting and make Zep's system distinct:

• Skills aren't tied to one model. A stronger model can solve a task once, and a weaker model can reuse the procedure without retraining.

• Zep uses the open Agent Skills format. Each Skill is packaged as a SKILL.md file with YAML frontmatter and optional files, so compatible harnesses like Claude Code and Codex can load it.

• Every Skill field needs evidence. The compiler links instructions to specific Trace events, making it possible to inspect where they came from. It can also patch individual fields instead of rewriting the whole Skill.

• Skills have a lifecycle. They’re versioned, checked before activation, and ranked partly by verified outcomes. A Skill can even start from one verified success, but remains provisional until it gathers more evidence.

The results are also fascinating:
- Skills learned from Opus 5.5 improved three weaker models' pass rates on the same tasks by 4.5–18.8 percentage points across 3 benchmarks.
- On EnterpriseOps-Gym, all three weaker models using Skills at low reasoning effort beat max effort without them.

Most gains came from tasks the stronger model had already solved.

Try Zep Agent Skills yourself here, you may find something interesting: getzep.com/zep-agent-skil…

How to replace a large, expensive, and slow model with a small, cheaper, fast model and still get the same results.

7x faster and 50% fewer tokens.

This is like fine-tuning, but without touching the model.

Here, we teach Haiku, using Zep Agent Skills, to do what Opus 5.5 does It takes 7 minutes with Opus 5.5. Only 59 seconds with Haiku.
• It takes 1.4M tokens with Opus 5.5. Only 700K with Haiku.

With Zep Agent Skills, DeepSeek V4.1 Flash passed more EnterpriseOps-Gym tasks than Claude Opus 5.5 at maximum effort.

Skills are procedures @zep_ai learns from your agent's traces. Each subsequent run can improve them.

Live today and free for 60 days.

It works with any model and framework. We import the traces you already send to Braintrust or LangSmith, and your agents read Skills over MCP or our SDK.

Agent Skills turns your agent's verified runs into procedures it reuses, so smaller models perform at or near frontier level for less. It's live today and free for 60 days: getzep.com/blog/introduci…

An agent that trusts the newest fact can get questions about the past wrong.

Every fact in a Zep Context Graph has two timelines: when it was true, and when the graph learned it. When a fact changes, the old one gets an end date and stays in the graph.

Full write-up below.

We read 15 write-ups from teams that built internal data agents, including OpenAI, Anthropic and LinkedIn.

13 of 15 rely on people to write or approve business context by hand. None describes recording when a business fact was true.

Full write-up below.

Vector RAG only gives agents reliable context when one chunk holds the answer.

It breaks down when answers span data sources or facts change over time.

A temporal context graph connects facts across sources and records when each was true. Full article linked below.

When an agent answers from a context graph, it's hard to see which facts it used.

Our graph assistant in the Zep dashboard shows each search it runs and the evidence behind the answer.

We launched Zep v4 this week with a lot of new features. See them all below.

Agents can make poor decisions with simple facts as context.

"Who sells part P-7 to Globex?"

A context graph with pairwise edges can't reliably answer that.

Our v4 API adds hyperedges to Zep Context Graphs, which give agents full facts.

Full write-up below.

Thank you to everyone who came out to FRC #26 at @pebble_bed for a16z Tech Week. Agents as database users turned into a packed room and a great debate. 🙏

Talks from @danielchalef (Zep), @shulynnliu and @shagarw21 (UC Berkeley Sky). Huge thanks to @Kristopherfloyd and izing.

Next one soon 👀

The Context API for Agents! 🔥

Alongside v4, we're announcing tons of new stuff over the next few days. Stay tuned!

Congrat to the @zep_ai team! 💙

Memory layer or a whole new database? We spent FRC #26 arguing about it 🧠🗄️

@danielchalef (@zep_ai) made the case that context is production infrastructure: if your agent can't trust what it remembers, nothing else matters.

@shulynnliu and @shagarw21 brought their CIDR '26 . Agents don't really query, they speculate: tons of redundant, exploratory probes. So build the database for the agent first.

Then the room fought over whether those are the same problem. No consensus. Exactly how we like it.

Huge thanks to @constructive_io and @danlynch for sponsoring and @pebble_bed + @nicilevv for hosting 🙏

Great to have @tjskhot @yumi_aw @ryanorban @ezrwolf @damnventures and many other regulars in the room!

Tonight: Frontier Research Club #26 at @pebble_bed, an official @Techweek_ by @a16z SF session.

Agents in Production: The Data Layer

@danielchalef of @zep_ai on context as production infrastructure: how facts go stale, how provenance and access control actually work, and how tale or cross-tenant context.

@shulynnliu and @shagarw21 of @UCBerkeley's Sky Lab on "Supporting Our AI Overlords," their paper with Ion Stoica and Matei Zaharia arguing agents don't query like humans, they speculate, so data systems need a rebuild.

A production memory system, then the case for a new kind of database. The room decides if they're describing the same thing.

Hosted by Kristopher Floyd and Dan Lynch, supported by @constructive_io. We're at capacity, waitlist below. Doors 5:30.

Terrible developer experience. @FireworksAI_HQ doesn't allow you to download models and LoRA adapters trained on their platform, despite documenting this capability. You have to ask for permission and wait several days! This despite spending $$$

🫤 Gating key functionality...

Our graph database at @zep_ai kept failing under load, and adding hardware stopped helping. We replaced it with one we built ourselves. The first version was my holiday project. 🧵

Coding agents built much of Konig, used adversarially: one agent proposes a design, another attacks it. The instrumentation came first, and agents could deploy a change, run the tests, read the traces, and fix what they found.

Cutting the scope made it buildable: no query language, and none of the surface area a general-purpose database accumulates. A small team built it and maintains it.

Full write-up: blog.getzep.com/why-we-built-a…