Overview
Lamis Mukta, Member of Technical Staff on Anthropic's Applied AI team, argues that raw model intelligence does not compound on its own. Agents need task-specific context that the newest model will not invent from scratch. Context engineering is the multiplier.
The one-year path: CLAUDE.md files that proved unreasonably effective, then in-band memory tools, then Skills with progressive disclosure, then memory modeled as a searchable markdown filesystem using ordinary bash and grep.
Production breaks the neat theory. Concurrent writes, stale or injected memories, and mixed human-agent edits require versioning, hash-checked concurrency, permissioning, and a portable API.
In-band memory still splits focus and cannot see cross-session patterns. Dreaming is the out-of-band batch process: transcripts plus the memory store go to a dedicated agent that proposes a cleaned store for human accept or reject.
Model releases keep arriving. Organizational knowledge does not. The durable advantage is a memory system that gets better the second time a task runs, drops token spend, and frees builders to ship product instead of re-prompting.
The talk is not coding-only. Mukta uses the same memory loop for slide style and writing preferences. The production pattern -- simple files first, then guardrails, then dreaming -- is the current Anthropic applied stack for long-running multi-agent work.
Key Points
- Model intelligence alone does not compound. Agents need organization-specific context that is often orthogonal to the model weights.
- CLAUDE.md files injected at session start were unreasonably effective for codebase navigation and user preferences -- until they bloated.
- Keep memory human-readable markdown that both humans and agents can write.
- In-band memory tools let agents decide when to read, write, and update during a session. Autonomy works.
- Skills solve ever-growing context via progressive disclosure: scan short front matter, load the body only when needed.
- State of the art: treat memory as a filesystem. Let agents use bash and grep instead of custom memory APIs.
- Production requires versioning (who, which session, rollback), hash-checked concurrency, permission layers, and a portable API.
- Do not let a single agent write organization-wide context. Scratchpads write; org memory is read-only unless gated.
- In-band memory splits tokens between the current task and future-self curation, and cannot see cross-session or fleet-wide patterns.
- Dreaming is an out-of-band batch job: memory store plus transcripts in, proposed store with examples and prevalence stats out.
- Effective memory raises second-run accuracy and cuts tokens and latency because agents one-shot more work.
- Start with the simple thing that works. Add guardrails only when you scale to many agents, long time horizons, or complex domains.
Quotable
Quotable moments are auto-generated from the transcript. Speaker attribution and quote accuracy should be verified against the original source before republishing or sharing.
Lamis Mukta -- Anthropic
"The intelligence alone is not going to compound because they need this context that helps them perform the specific tasks that you need them to."
Why it works: separates model IQ from usable product performance in one sentence.
Lamis Mukta -- Anthropic
"At Anthropic we like to say do the simple thing that works."
Why it works: licenses CLAUDE.md and markdown files as first-class architecture, not a hack.
Lamis Mukta -- Anthropic
"Agents are actually just very good at using normal file system tools like Bash and grep. So just let them search over the file system rather than being opinionated about the specific tools."
Why it works: kills the urge to invent a custom memory API before agents have proven they need one.
Lamis Mukta -- Anthropic
"You are asking an agent to complete a task but at the same time you are also asking it to invest in memory curation which would help it perform better in a future run."
Why it works: names the in-band split of focus that dreaming is designed to remove.
Lamis Mukta -- Anthropic
"Keep thinking, keep learning, and keep dreaming."
Why it works: closes the talk by turning the product name into the operating stance.
Concepts
Context primitives
Context as multiplier
Newer models do not know your org, codebase, or taste. Investing in context multiplies intelligence as models improve instead of resetting the learning curve on every release.
CLAUDE.md
A short markdown briefing loaded at session start. Unreasonably effective until it grows long enough to bloat the window. Still the correct first move.
Progressive disclosure (Skills)
The agent scans a few sentences of front matter the way you scan book titles on a shelf, then loads the full procedure only when relevant. Deep detail without permanent context tax.
Memory as filesystem
Store memories as markdown in directories. Index them. Let agents search with ordinary tools. Do not invent a special memory CRUD layer until the filesystem fails.
Production guardrails
Versioning
Keep versions, the originating session or transcript, and the human or agent author so you can roll back a bad write.
Hash-checked concurrency
Hash before draft, hash before commit. On mismatch, re-pull, redraft, retry. Stops two agents from silently overwriting the same file.
Permissioning
Org-wide principles stay curated and mostly read-only. Team slices sit in the middle. Agent scratchpads are writable. One bad org write scales to the whole fleet.
Portability
Curated memory is an asset. Expose it through a clean API so more than one product surface can read it.
Learning loop
In-band memory
Read and write inside the same session that is doing the user's task. Fast feedback to the next run. Limited visibility and competing token budget.
Dreaming
Out-of-band batch process with dedicated tokens. Reviews transcripts (including tool-call metadata), finds fleet-wide patterns, proposes a new memory store with examples and prevalence stats. Humans accept or reject.
School analogy
Students (session agents) submit work. Teachers mark it. A head teacher sees the whole fleet and rewrites the curriculum. Dreaming is the head teacher.
Harness vs autonomy
Let agents write markdown freely until the primitives stabilize. Then bake versioning and hashing into the harness so those parts stay deterministic.
Implementation
Implementation steps are auto-generated from the transcript content and are provided for informational purposes only. They do not constitute professional advice of any kind. Always consult a qualified professional before acting on any information presented here.
Start with a short CLAUDE.md
Write a human-readable markdown file covering codebase map, org norms, and personal preferences. Load it at session start. Keep it short. Do not solve scale problems you do not have yet.
Give the agent write access to memory files
Let the agent decide when to read, write, and update during a session. Autonomy here is a feature. Keep the format markdown so you can audit it.
Move procedures into Skills
For end-to-end workflows you care about, put a short description in front matter and the full procedure in the body. The agent loads depth only when the title matches the task.
Treat memory as a searchable filesystem
Organize markdown in directories. Index it. Let agents use bash and grep. Do not invent a custom memory tool until search over files fails.
Add production guardrails
Version every write with author and source session. Hash before commit and retry on conflict. Split org-wide (read-only) from scratchpads (write). Expose memory through a portable API.
Keep in-band writes small
Session agents should jot what will help the next run of the same task. Do not ask them to reorganize the whole store while they are also doing the job.
Run dreaming out of band
On a schedule, feed the current store plus a permission-matched set of transcripts (dialogue and tool metadata) to an orchestrator and sub-agents. Require proposed diffs with examples and prevalence stats. Accept or reject by hand at first.
Steer what dreaming is allowed to care about
Tell the dreaming agents which patterns matter for your org and which noise to ignore. Mirror live-agent permission sets when you select transcripts so out-of-band jobs cannot leak context across tenants.
Tools & Resources
These resources are curated in two groups. Mentioned Resources are pulled directly from the source material, and Suggested Resources are added to help you expand and apply the ideas beyond the original.
The following resources may contain affiliate links. As an Amazon Associate I earn from qualifying purchases at no extra cost to you. This does not influence the placement of links on this page.
Mentioned
| Resource | Notes |
|---|---|
| CLAUDE.md / Claude Code memory | Official hierarchy for project, user, local, and enterprise memory files. |
| Agent Skills | SKILL.md format, front matter, progressive disclosure. |
| Steering Claude Code | When to use CLAUDE.md vs Skills vs hooks vs subagents. |
| Managed Agents Dreams API | Out-of-band dreaming: memory store + sessions in, new store out. Input store is not mutated. |
| Dreaming product announcement | How Anthropic positions memory plus dreaming for self-improving agents. |
| Talk video (AI Native DevCon) | Lamis Mukta on context engineering, memory, and dreaming. |
| Source clip on X | Posted by Dhruv kumar; talk by Lamis Mukta. |
Suggested
| Resource | Notes |
|---|---|
| Designing Data-Intensive Applications -- Kleppmann | Versioning, concurrency, and hashing are database problems. Mukta says the harness should absorb them once the primitives are known. |
| Building Microservices -- Newman | Permission boundaries and portable APIs for shared organizational context. |
| Claude Code memory docs | Start here if you are implementing the simple thing that works on a laptop first. |
| The New Stack on dreaming | Independent writeup of the same Mukta talk, including the injection and staleness risks. |
AI Prompt
Implementation prompt