Video summary
Anthropic: Context Engineering & Autonomous Memory for AI Agents | Lamis Mukta (AI DevCon)
Main summary
Key takeaways
Summary
Anthropic’s Lamis Mukta describes context engineering as a way to make model intelligence useful and cumulative in real deployments. A capable model does not automatically know an organization’s codebase, workflows, or preferences. Supplying and maintaining that context can improve task performance over time.
How agent memory has evolved
CLAUDE.mdfiles: Simple, human-readable Markdown instructions can help agents understand a codebase, an organization, or user preferences. They work well, but can grow too large if everything is loaded into the context at the start of every session.- Memory tools: Letting agents decide when to read, write, or update memories can support learning during a session. This is “in-band” memory management because it happens within the task session.
- Skills and progressive disclosure: Skills are suited to procedural workflows. An agent can inspect a short description first and load a more detailed skill only when relevant, preserving context space while still making detailed guidance available.
- File-system-based memory: Mukta’s current recommended approach is to store memories in searchable files, often in Markdown, and let agents use familiar tools such as
grepto find what they need. The key is to allow memories to grow while making relevant information easy to retrieve, and to give agents autonomy in writing them.
Production considerations
Simple memory approaches become harder to manage when many agents and people share a workspace. Mukta highlights several safeguards for production systems:
- Versioning: Track changes, their authors, and the sessions or transcripts behind them, and allow rollback.
- Concurrency control: Use hashes to detect whether a memory changed while an agent was drafting an update. If it changed, the agent should reload the latest version and retry.
- Permissions: Separate organization-wide, shared, and individual working memories. For example, broad organizational context may be read-only for agents, while an agent can write to its own scratchpad.
- Portability: Keep memory accessible across products and systems through a clean interface or API.
- Staleness and integrity: Memories can become outdated, incorrect, or maliciously altered, so production systems need ways to review and manage them.
Mukta says effective memory can help agents do tasks more accurately on later attempts, use fewer tokens, and complete work faster. She also stresses that these principles apply beyond coding, such as retaining presentation or writing preferences.
Out-of-band memory curation: “Dreaming”
In-band memory has two limitations: an agent must divide attention between its current task and future learning, and it sees only its own session, so it may miss patterns across sessions or agent fleets.
Mukta presents dreaming as an asynchronous, out-of-band process that uses dedicated resources to review memory stores and agent activity. It can examine transcripts—including tool calls and related metadata—look for recurring failures or missing guidance, and propose memory updates.
An orchestrator can collect analyses from sub-agents and present proposed changes, with examples and prevalence information, for a person to accept or reject. The process can be tailored to an organization and configured to use transcripts consistent with the relevant memory permissions.
Practical guidance and Q&A
The talk’s main recommendations are to start with simple tools such as Markdown files and skills, add production safeguards as systems scale, and consider a separate curation process to consolidate useful memories and remove stale ones.
In response to an audience question about enterprise storage, Mukta pointed to Anthropic’s memory and dreaming API for managed agents as an out-of-the-box option. She also explained that versioning and concurrency controls are increasingly being codified in the agent harness as deterministic software-engineering practices, rather than left entirely to agent discretion.
Main speaker/source: Lamis Mukta, Technical Staff on Anthropic’s Applied AI team. Other sources: An event moderator and audience members during the Q&A.
Rate this summary
Your feedback will help improve summaries.
Improve this summary
Reprocess with a stronger model when the summary feels incomplete or inaccurate.
Translate summary in another language
Ask questions to this video
Chat for follow-up questions, clarifications, and source-backed answers.