[[Tim Kellogg]] notices that three recent posts on [[AI agent|agent]] [[Context management|context management]] - Anthropic's, Cognition's, and Letta's - are all saying the same thing: to remember is to compress, and compression is [[Cognition|cognition]]. ## Layered Memory as Lossy Compression - [[Lossy compression]] - Letta tiers an [[AI agent|agent]]'s memory into four blocks (core, message buffer, archival, recall), each a lossier compression of the last, like L1/L2/L3 CPU cache. - The [[AI agent|agent]] promotes details up to long-term storage and rehydrates them on [[Query|query]] - a cascade of compressions and selective decompressions. - [[Rate-distortion tradeoff]] - Kellogg reads the memory hierarchy as this classic information-theory curve: > **Rate–distortion tradeoff:** This hierarchy embodies a classic principle from information theory. With a fixed channel (context window) size, maximizing information fidelity means balancing **rate** (how many [[Token|tokens]] we include) against **distortion** (how much detail we lose). ^kellogg-rate-distortion - The payoff line of the section - managing memory is itself an act of [[Cognition|cognition]], because choosing the gist and forming the [[Query|query]] is understanding: > This layered approach turns memory management into an act of **cognition**. ^kellogg-memory-is-cognition ## One Mind vs. Many Minds: Two Approaches to Compression - [[Multi-agent]] - the alternative to one mind compressing over time is many minds compressing in parallel, each its own [[Context window|context window]] - the diverge step of [[Folding context]]; Kellogg leans on Anthropic for the claim that search itself is compression. ![[How we built our multi-agent research system#^search-is-compression]] - The cost - Kellogg turns to Cognition's [[Walden Yan]], who argues such systems are fragile because each [[AI agent|agent]] compresses its slice in isolation with no [[Context management|shared context]]: ![[Don't Build Multi-Agents#^too-dispersed]] - Cognition's answer is not to fork minds but to compress harder in one - a dedicated model condensing the history into key details, events, and decisions: > if you must lose information, **lose it intentionally** and in one place - via a trained compressor - rather than losing it implicitly across multiple agents' blind spots. ^kellogg-lose-intentionally ## Conclusion: Compression is Cognition - The convergence - intelligence is bounded by information [[Bottleneck|bottlenecks]], and overcoming them looks like compression; Kellogg summarizes what [[Cognition|cognition]] therefore is: > An effective mind (machine or human) can't and shouldn't hold every detail in working memory – it must aggressively **filter, abstract, and encode** [[Information|information]], yet be ready to recover the right detail at the right time. ^tim-kellogg - The closing image - to extend an [[AI agent|agent]]'s [[Cognition|cognition]] is to engineer the art of forgetting; memory is meaningful precisely because it is prioritized, lossy, and alive.