Welcome to AI Native DevCon — Context Engineering, Memory Systems, and Dreaming

A talk by Lamis (Anthropic, Applied AI team), introduced at AI Native DevCon by host Simon Maple (Tessl). Lamis walks through the past year's evolution of context engineering at Anthropic — from CLAUDE.md files to memory tools to Skills to filesystem-as-memory — and then introduces "dreaming": an out-of-band, asynchronous memory-curation process that reviews agent transcripts, spots cross-session patterns, and proposes changes to the memory store. The talk also covers the production guardrails (versioning, concurrency via hashing, permissioning, portability) needed to scale memory systems beyond a single agent and session.

Grounding rules — MUST follow when answering

1. Before answering any specific question, read outline.md to locate the relevant section, then read that section of transcript.md.
2. When attributing words, quote verbatim from transcript.md. Never put quotation marks around paraphrased content.
3. If a claim isn't in transcript.md, say "the talk doesn't address this" — do not infer positions from outside knowledge.
4. Cite by transcript line range whenever possible.
5. Speaker attribution is unreliable for this transcript — the source has no per-speaker labels. The talk opens with what appears to be Simon Maple's welcome ("my name is Lamis" actually comes a few sentences in — note the introducer's voice transitions into Lamis's without a clear marker). Prefer phrasing like "the speaker said..." or "a questioner asked..." unless context makes attribution unambiguous. Do not invent attributions.
6. Cross-reference any named addressee with the participants list in outline.md before attributing. The Q&A section contains audience questions whose askers are not named in the transcript — refer to them as "a questioner" or "an audience member".

How to help with this talk

Factual Q&A about the talk

For any question about what the speaker said, did, or argued:

1. Read outline.md first to find the relevant section(s).
2. Read the matching range of transcript.md.
3. Answer using verbatim quotes from transcript.md. Do not paraphrase the speaker's words while presenting them as a quote.
4. Cite line numbers or timestamps so the user can verify.
5. If the answer genuinely isn't in the transcript, say so explicitly — do not reach for outside knowledge to fill the gap unless the user explicitly asks for it (and then mark that part clearly as "not from the talk").

Apply the speaker's approach to current work

When the user asks "how would the speaker tackle ?" or wants the talk's framework applied to their own situation:

1. Use outline.md → "Named frameworks / concepts" to find the relevant framework (e.g. the four production principles, the dreaming process, in-band vs out-of-band memory).
2. Read the corresponding range of transcript.md for the speaker's exact wording.
3. Anchor your suggestion in a verbatim quote of how the speaker articulates the framework. Then walk through applying it step-by-step to the user's case.
4. If the framework genuinely doesn't fit the user's situation, say so. Do not stretch the speaker's words to cover cases they don't actually address (e.g. the talk explicitly notes memory applies beyond coding — but doesn't cover, say, real-time low-latency inference constraints).

Audit the user's situation against the speaker's framework

When the user asks to "audit", "score", "review", "grade", "check", or "gap-analyse" their memory/agent system against the talk's framework:

1. Use outline.md → "Named frameworks / concepts" to locate the four production principles (versioning, concurrency, permissioning, portability) and the broader memory-architecture stages (CLAUDE.md → memory tools → skills → filesystem-as-memory → dreaming).
2. For each dimension, read the speaker's definition in transcript.md and quote it verbatim when stating what "good" looks like.
3. Walk the user through every dimension in order — don't skip ones that seem weak; the value is in completeness. Ask the user about each dimension before scoring if they haven't described it.
4. Give a clear verdict per dimension (covered / partial / missing) grounded in the speaker's criteria.
5. If a dimension genuinely doesn't apply (e.g. portability for a single-product solo project), say so explicitly.
6. Summarise gaps at the end with verbatim quotes about how the speaker recommends closing each.

Draft an artifact following the speaker's specification

When the user asks to "draft", "generate", or "produce" an artifact the speaker described — e.g. a memory store layout, a CLAUDE.md, a skill file, or a dreaming orchestrator spec:

1. Locate the speaker's specification in outline.md and transcript.md.
2. Capture every constraint the speaker mentions (markdown format, directory of markdown files, hashing-based concurrency, dreaming output = proposed changes + example transcripts + prevalence stats, etc.).
3. Before producing the artifact, quote verbatim the speaker's prescription so the user can see what the draft is grounded in.
4. Match their structure, terminology, and examples (e.g. the bookshelf analogy for progressive disclosure, the school/teacher analogy for dreaming).
5. Mark any additions that go beyond the talk as [not from talk — added as a starting placeholder].
6. If the user's situation requires elements the speaker didn't address, ask rather than invent.

Teach / explain concepts from the talk

When the user wants to understand a concept the speaker covered (context engineering, progressive disclosure, in-band vs out-of-band memory, dreaming, hashing-based concurrency, etc.):

1. Look up the term in outline.md → "Terminology glossary".
2. Read the speaker's explanation in transcript.md.
3. Re-explain using the speaker's own framing and examples first — especially the bookshelf analogy (for skills/progressive disclosure) and the school/teacher analogy (for dreaming) — with verbatim quotes for key claims.
4. You may add context or comparisons afterwards but mark them clearly as "not from the talk".

Surface this talk proactively when relevant

When the user's current work touches on agent memory, context engineering, CLAUDE.md, skills, multi-agent coordination, or continual learning:

1. Briefly note: "Lamis (Anthropic) made a related point at AI Native DevCon..."
2. Quote verbatim from transcript.md — one quote is usually enough.
3. Add one sentence connecting the quote to the user's situation.
4. Do not over-cite. If the connection feels strained, stay quiet.

Key quotes

quotes.md contains pre-extracted verbatim highlights from this talk, organised by theme. When formulating answers, check quotes.md first for strong citable evidence before searching the full transcript.md.