Outline — Context Engineering, Memory Systems, and Dreaming — Lamis

Speaker

Lamis — Member of Technical Staff at Anthropic, on the Applied AI team. Per the talk: "this is a team which sits between research, product and go-to-market. So we do a mixture of working on internal projects as well as directly with customers who are building agents at the frontier." Lamis works specifically with startups and founders.

Introducer / host: Simon Maple, Head of Developer Relations at Tessl and AI Native Dev co-host. Previously Field CTO and VP Developer Relations at Snyk, ZeroTurnaround, and IBM. Java Champion (2014), JavaOne Rockstar (2014, 2017), Duke's Choice award winner, Virtual JUG founder.

Note: the transcript has no speaker labels. The opening welcome ("Hey everyone... hopefully we'll be having a wonderful day so far here at AI DevCon") reads as Simon's host intro, but transitions into Lamis's talk by the line "By way of introduction, my name is Lamis." without a clear break.

Abstract

[Not provided by the user.] [Inferred:] A tour of the past year of context-engineering primitives at Anthropic — CLAUDE.md, memory tools, Skills, and filesystem-as-memory — followed by the production engineering guardrails (versioning, concurrency, permissioning, portability) needed to scale these to multi-agent fleets, and culminating in dreaming: an out-of-band, asynchronous memory-curation process inspired by the analogy of a head teacher reviewing student work to spot patterns and update the curriculum.

Thesis (synthesis)

Raw model intelligence doesn't compound in your organization unless you invest in context engineering — and the state of the art has converged on filesystem-style memory stores of markdown files that agents autonomously read, write, and search. But to scale these to production you need classic engineering primitives (versioning, hash-based concurrency, permissioning, portability) baked into the harness. And to escape the limits of in-band memory (where the same agent juggles task-completion and memory curation), you add an out-of-band asynchronous process — dreaming — that reviews many transcripts at once, finds cross-session patterns, and proposes memory updates.

Section TOC

Terminology glossary (definitions the speaker actually gave)

• Context engineering — the work of translating raw model intelligence into "durable, scalable, useful product" by giving agents the context they need to perform org-specific tasks. "a really great investment ... has the effect of multiplying the intelligence even as models get smarter."
• CLAUDE.md — a markdown file shipped with Claude Code that "just gives the agent a couple of instructions about making your way around the code base, the organization, your own user preferences" and is injected at the beginning of a session.
• Memory tools — tools that let agents "autonomously manage their own memory systems ... decide when they read, when they write and when they update memories ... all happening in band."
• Skills — units with a few-sentence frontmatter the agent scans first, plus a deep body loaded only when relevant. Suited to "processes where you have like a procedural workflow ... an opinion about how the process works end to end."
• Progressive disclosure — the design idea behind skills: surface a short title/frontmatter for discovery, only load full detail when needed. Speaker's analogy: a bookshelf of books you scan by title and pull out when relevant — "if someone walked up to me and started speaking to me in French and I noticed I have a French dictionary, I can pull that out."
• Filesystem-as-memory — current state of the art per the speaker: "modeling these memory systems just as file systems ... fill them up with markdown. Agents actually just very good at using normal file system tools like bash and rep."
• In-band memory — memory read/written by an agent within a specific session. Limitations: "inherent split of focus and resources" between task and curation, plus per-session visibility only.
• Out-of-band memory curation — a separate asynchronous process with its own resources and broader visibility, dedicated to memory hygiene.
• Dreaming — speaker's term for an out-of-band process: "a second order process over memory ... runs in batch and asynchronously, with its own allocated resources to ensure that those memories themselves are affected up to date." Inputs: existing memory store + a batch of agent transcripts. Output: proposed changes to the memory store, with example transcripts and prevalence stats.
• Memory store — "a collection of memories. Memories themselves might just be marked down files. Organized in this directory."
• Steering (of dreaming) — telling the orchestrator/sub-agents "in your specific case, these are the kinds of things I think are important and relevant. These are the kinds of things that are not important and relevant."

Named frameworks / concepts

The four production principles for memory systems

1. Versioning — store versions, allow rollback, track what context the update was based on (which session, which transcript) and who/which agent made it.
2. Concurrency (hash-based) — before writing, the agent hashes the memory; drafts its edit; hashes again before commit. "If those two things do not match then the agent cannot write it because it means that some update was made in the meantime." Agent then re-drafts.
3. Permissioning — tier memory from "top level organizational wide knowledge" (often read-only) down to per-agent scratch pads (read/write). Guardrails prevent one agent corrupting org-wide context.
4. Portability — clean API so the curated memory is accessible across multiple products and systems.

The context-engineering timeline (Anthropic's "do the simple thing that works")

CLAUDE.md → autonomous memory tools → Skills (progressive disclosure) → Filesystem-as-memory → (production guardrails) → Dreaming (out-of-band curation).

Dreaming process

• Inputs: existing memory store + batch of transcripts (including tool calls and metadata, not just text exchanges)
• Orchestrator deploys a fleet of sub-agents to analyze transcripts
• Steering: humans tell the system what's important/not
• Orchestrator aggregates patterns, proposes memory changes
• Output includes example transcripts demonstrating the pattern and prevalence stats justifying the change
• Human reviews and accepts/rejects

Two analogies the speaker uses

• Bookshelf — for progressive disclosure / skills.
• School (students, teachers, head teacher) — for dreaming: dedicated capacity for helping people learn, plus fleet-wide visibility to spot patterns and update the curriculum.

Open questions / not covered

• The talk does not give a concrete spec for the markdown directory layout of a memory store beyond "marked down files. Organized in this directory."
• It does not benchmark dreaming quantitatively — claims of "better accuracy", lower cost, lower latency are stated as observations without numbers.
• It does not address adversarial robustness in detail beyond mentioning "maliciously injected" memories as a risk.
• It does not compare dreaming to existing eval-loop or RLHF practices.
• It deliberately avoids product-pitching until directly asked in Q&A ("I was kind of coy in the talk because we're not allowed to make product transactions"), at which point it points to Anthropic's memory and dreaming APIs via Claude Managed Agents.
• The Q&A on permissions across enterprise users gets a high-level answer ("you decide exactly which session transcripts to attach") but does not detail concrete mechanisms.
• The "reinventing databases" question gets a philosophical answer ("threading the needle") but no concrete guidance on when to use a real database vs the markdown approach.
• The talk does not cover real-time / low-latency inference constraints, cost ceilings for dreaming, or how often dreaming should run.