Consolidation and co-evolution

On September 21, 2026, Ye, Lu, Dong, Su and Song posted Harness-Zero, a recipe for training a harness into a model and then throwing the harness away ¹. They evolved a specialized harness for each domain (knowledge work, tool use, science) using Kimi K3 under Kimi Code. They then ran a Qwen3.5-9B student inside a deliberately bare target harness, a minimal mini-SWE-agent-style loop with a fixed system prompt and one Bash tool, while GPT-5.6 Sol, guided by the evolved harness, corrected the student's responses before they executed. Fine-tuning on those corrected trajectories produced a model that, with the specialized harness removed, raised macro-average task success from 23.3% to 44.3%. With the specialized harness still attached, the untrained base model reached 41.7%. The distilled model recovered, on average, 82.3% of 28 harness-induced behavior patterns that the base model lacked ¹.

The result is a week-old preprint on one student model, and the teacher in the loop is a frontier model, so part of what got distilled is GPT-5.6 Sol's judgment rather than the harness alone. The result still makes the storage question concrete. An evolved harness is expensive to build and tied to one domain; a general-purpose agent either settles for a mediocre shared harness or routes among a growing zoo of specialized ones. Harness-Zero's answer is to use the specialized harness as training-time guidance, write what it induces into the weights, and ship one plain harness. The harness did the discovery; the weights did the keeping.

Every self-improving system has to decide where its improvements are written. Written into text (prompts, memory, skills, harness code), learning is cheap, legible, reversible, and available on closed models. Written into weights, it is compact, fast at inference, and able to change what the model can do rather than what it is told. The hard question is no longer which store wins. The 2026 evidence is about the machinery that uses both, moves lessons from one to the other, and decides which lessons belong where.

The sharpest statement of the context-first position comes from Letta, which builds stateful agents. In June 2026 it wrote: "While models are a depreciating asset, learning and memory represented in token space is an appreciating asset" ², sharpening its December 2025 line: "The weights are temporary; the learned context is what persists" ³. The reasoning is practical. Most builders use closed models whose weights they cannot touch, a new model generation arrives every few months, and a memory file or skill library carries over to the next model while a fine-tune does not. Letta's roadmap calls for "memory models" trained with RL to write durable context "that transfers across model generations" ² (see 07, Experiential memory).

The consolidation camp argues that context is a staging area. In May 2026 Dwarkesh Patel, relaying Andrej Karpathy, put it this way: "Our working memory gets wiped regularly. What we actually have is a consolidation process (sleep) that distills stuff into the brain, in a weird and lossy way… It's not obvious it's something you can get for free from doing long enough RL loops" ⁴. Within eight days two papers put sleep in their titles. One converts recent context into persistent state-space fast weights through offline recurrent passes and then clears the KV cache ⁵; the other adds a "sleep" stage in which a model consolidates memories by distillation and "dreams" an RL-generated curriculum to rehearse new knowledge ⁶.

Both camps now concede a hybrid. Letta grants that "the primary advantage of parametric memory is efficiency" and that distilling memories into weights "can provide additional gains in personalization and efficiency" when the weights are accessible ^{3 2}. The sleep papers treat context as the place where new information lands first. The disagreement is over which store is primary and how often the slow one should be written.

Complementary Learning Systems (CLS) theory gives the structure a name. McClelland, McNaughton and O'Reilly argued in 1995 that the brain needs a fast hippocampal store that records episodes after one exposure and a slow neocortical store that learns structure through interleaved replay, because one store cannot do both without catastrophic interference ⁷. The mapping to agents is direct. Context, memory, skills and harness are the fast store: written every episode, specific, cheap to revise. Weights are the slow store: updated rarely, general, expensive to revise, and prone to interference when updated carelessly. Consolidation is replay from the first into the second. The analogy's one testable prediction is that the slow store should learn from what the fast store computes, not from raw experience; a 2025 DeepMind study co-authored by McClelland found that fine-tuning generalized better when its data included the model's in-context reasoning traces ⁸. By August 2026 engineers were building CLS into agent memory by name: Dual-Layer Agentic Memory routes incoming facts into an external store and periodically consolidates high-value entries into the weights with supervised fine-tuning; a 1.7B/8B routing cascade pruned up to 68% of redundant external memory while keeping over 98% of the QA exact match of a keep-everything baseline ⁹. That is enough work for the analogy. The engineering question is how a loop should switch between the two stores.

The first systems to put both levers inside one loop arrived between May and September 2026, and they share a shape. The harness is edited while weights are held fixed, the weights are trained under the improved harness, and each change exposes a new bottleneck in the other. What changed over the summer is the answer to how often to switch.

WHALE, posted August 31 by Kim, Lee, Finn and Lee, is the cleanest test ¹⁰. It alternates online rejection-sampling fine-tuning with Meta-Harness search (Claude Opus 4.7 as the harness proposer), using Qwen3.5-2B and 4B agents on search QA, math and chess puzzles. It beat weight-only, harness-only and Fast-Slow Training baselines by 4.15 to 24.38 points in best mean@8 accuracy. The per-domain detail matters most: "Either component can be the bottleneck: harness search matches peak weight-only accuracy with far fewer rollouts in SearchQA, but improves math accuracy only after a weight update." And small interleaved updates beat a stagewise schedule that completes all weight optimization before beginning harness search, by 5.32 and 9.16 points in the two domains the authors report, at lower rollout cost ¹⁰. At this scale the answer to "harness or weights first" is neither: switch often.

Earlier systems switched once or twice. SIA, from May, lets a Claude Sonnet 4.6 Feedback-Agent choose after each generation between rewriting a gpt-oss-120b agent's scaffold and running RL on its LoRA adapter; in practice it ran harness search until progress stalled and then switched ¹¹. On LawBench's 191-class charge prediction, top-1 accuracy went from 13.5% to 50.0% after harness search to 70.1% after weight updates, against a prior best of 45.0%. On a single-cell denoising task, scaffold rewrites carried the score from 0.048 to 0.241 and stalled; the first weight-update checkpoint then wrote a two-line np.clip plus np.rint step that rounds imputed counts to non-negative integers, "a biological invariant that is trivially correct yet absent from any prior scaffold version," lifting the score to 0.289 ¹¹.

Co-Harness, from July, adds explicit acceptance rules ¹². An LLM "HarnessCritic" labels each failed trajectory with a root cause and proposes a local diff, which is kept only if it improves a held-in slice showing the failure without regressing a held-out slice; accepted diffs go into a versioned registry "that serves as an audit trail and enables full rollback." Verifier-passing trajectories under the accepted harness then fine-tune the next model. On tool-integrated math with Qwen3-8B and Qwen3-32B, average pass@1 across AIME24, AIME25 and HMMT25 moved from 54.2 with a human harness to 58.5 with the evolved harness, then 73.5 and 78.9 after two rounds of fine-tuning. Its 200-hour case study shows what the harness lever buys: the initial configuration crashed after 32 seconds on a vLLM KV-cache misallocation, and over 22 versions the critic fixed that, cured a zombie-thread out-of-memory failure, and found an 8.7× batching speedup. It also added domain prompts that cost about 16.5 points by clashing with Qwen3's trained reasoning format, and the registry rolled them back. No amount of RL on a trainer that crashes in 32 seconds produces a gradient.

Other August systems make the harness an explicit data engine. HELIX, from HKU, ran one evolution round on code repair: a 65-candidate harness portfolio found a fixed harness that improved task coverage by 4.0% over the Pi agent, and a 200-slot slice of sibling runs yielded 438 verified SFT, critic, filter and preference records for the next model ¹³. Macaron-V1, Mind Lab's August technical report, is the production-scale design: each release is a versioned model–harness pair (a frozen base, four LoRA specialists selected per turn, and a TOML "Harness Context Protocol" runtime configuration), new capabilities appear first as harness artifacts, and "when trajectory evidence warrants a parameter update, the change can then be transferred into a specialist adapter" ¹⁴. Its authors are candid that "the present evaluation measures the configuration-search stage, not repeated transfer across adapter generations." EvoTrainer, from June, applies the idea to the RL training harness itself, reporting in its abstract that it matches or exceeds human-engineered RL recipes on math, competitive programming and repository-level SWE ¹⁵.

The mechanism is consistent across these papers. The harness fixes failures that are operational rather than cognitive, and those failures bound what any training signal can reach. Weight updates need decent rollouts to learn from, which the harness supplies. And the weights then reach behaviors the harness cannot express, which change which harness is best, which is why WHALE's frequent switching wins.

Once a loop decides a lesson should move into weights, the next question is how. Every method is a variant of context distillation, an operator dating to 2022: run the model with privileged context as a teacher, train it without that context as a student ¹⁶. The variants differ in who samples the trajectories and what the teacher sees, and in 2026 those differences turned out to be enormous.

Experience Distillation, a July 2026 preprint, gave a model long histories of its own prior attempts at a task (60 to 600 turns, more than 80k tokens) and measured how much of the in-context gain survives once the history is removed and the lesson trained in ¹⁷. A teacher conditioned on the experience generated one-step decisions at each recorded point, and the student was trained on them with teacher-sampled forward KL. That kept 64.8% of the in-context gain on 749 SWE tasks and 93.4% on six TaleSuite text games. Plain supervised fine-tuning on the same experience kept 3.8% and −2.6%. On two games, on-policy distillation with reverse KL, which much of the field treats as the better default (see 04, Self-training and distillation), kept 9.1% and 0.4% at matched compute, against 96.7% and 98.4% for forward KL, because "without the experience context, the student rarely samples trajectories containing the high-quality decisions." The caveat is scope: the paper measures retention on the same tasks the experience came from, not transfer to new ones, and the base models are in-house.

September's results locate the rule more precisely. Yu et al. first evolved a harness for a weaker model on seven enterprise agent tasks, then trained the weaker model on a stronger expert's complete trajectories under that harness ¹⁸. Performance regressed on all seven tasks by 4 to 30 points across Qwen3-Coder and Gemma 4, though the same imitation helped under the unevolved harness. The weaker model "adopts the expert's planning strategy without the competence to execute it and no longer matches the harness evolved around its native planning style." Their fix keeps the student's own rollout, localizes the failing turn, and asks the expert to rewrite only that turn. Harness-Zero's correction-before-execution works the same way ¹, as does Cursor's Composer 2.5, which in May began inserting a short hint at a turn where the model went wrong (for a hallucinated tool call, a "Reminder: Available tools…" note) and distilling the hinted distribution into the unhinted model for that turn alone ¹⁹. Taken with Experience Distillation, the lesson is that consolidation targets must be dense and informed by the context, but they must sit on states the student visits itself and match the harness it will run in. Raw imitation of someone else's trajectories fails the second test; student-sampled reverse KL can fail the first.

The bandwidth argument explains why distillation beats outcome RL for this job. Thinking Machines argued in 2025 that policy-gradient RL delivers "only O(1) bits per episode" while supervised targets carry information on every token ²⁰. A context-conditioned teacher turns a rich context into token-level targets that encode what the context changed about each decision; an outcome reward carries a bit or two.

Consolidation also needs a stopping rule, and in September one appeared. RetireOPD trains a skill-free student with RL plus on-policy distillation from a skill-conditioned teacher, and lets the student drop the teacher "once their discrepancy stops shrinking and it reaches a target fraction of the teacher's success rate," continuing with RL alone ²¹. Across Qwen2.5 models from 1.5B to 7B it improved ALFWorld success over an RL baseline by 14.1% to 18.8% and WebShop accuracy by 11.8% to 19.0%, and surpassed its own teacher in every setting. The authors also found that "privileged information alone does not always make a teacher reliable," so the teacher is itself trained against environment rewards first. Retiring the teacher is scaffold deletion at training time.

The last failure mode is overfitting to one harness. Alibaba's Taobao Live team found in August that fine-tuning a compact model under a fixed harness configuration lowered IFEval by 7.7 points; training instead under task-preserving transformations of skill identifiers and content, tool schemas, prompt structures and hook functions, followed by on-policy distillation and RL in the augmented environments, avoided the regression (83.5 on IFEval) and scored 94.6 on a harness-variant QA set against the base model's 75.4, in a system deployed with positive online A/B results ²². Consolidating a harness into weights should make a model better inside that harness without making it brittle outside it.

Two older results bracket these. SEAL (2025) let a model write its own fine-tuning data and showed both the promise of self-directed consolidation and its cost in forgetting under sequential edits ²³; Cartridges (2025) showed a trained KV cache can match in-context learning on a document at 38.6× less memory ²⁴. Neither changes the 2026 picture.

A hybrid only makes sense if the levers compose. If RL simply relearned what prompt optimization found, or overwrote it, the second step would be wasted work. The direct experiments in the DSPy lineage (see 06, Prompt and program optimization) find composition. mmGRPO, revised in May 2026, reported that averaged over three seeds, two 8B models and three tasks, vanilla chain-of-thought scored 66.3, MIPROv2 prompt optimization 70.0, multi-module GRPO alone 71.2, and prompt optimization followed by GRPO 73.4 ²⁵; the authors attribute the gain to "the value of high-quality rollouts at the start of mmGRPO training," and the prompt step cost 1.4 hours on one H100 against 18.7 hours on two for RL. E-SPL, from February, samples each RL batch under several evolving system prompts and moved an AIME-trained model's success on BeyondAIME from 38.8% with RL alone to 45.1%, where prompt evolution alone reached 40.0% ²⁶.

Until August, every composition result ran prompts before weights, and the reverse order (optimizing a harness for a model that has already been trained) was untested. WHALE is the first controlled look: its math harness gains appeared only after a weight update ¹⁰. That is on 2B models, and at frontier scale the question remains open, though it is the situation every lab is in when it ships a new model into an existing harness. Co-Harness's clashing domain prompts are the small-scale warning that a context instruction can fight a trained-in behavior ¹².

The pattern is consistent: harness-side edits deliver a fast step, weight updates reach what harness search does not, and the way the weight step is taken decides whether the harness's lessons survive. The weaknesses are just as consistent, and they concentrate in the systems that carry the headlines.

SIA's LawBench number is the clearest problem. The paper describes RL rollouts as "generated solution script[s] executed against the test split" while stating that "all evaluations are on the held-out test split" ¹¹. Read as written, the RL reward was computed on the reporting split, which would make 70.1% an optimization-on-test number; SIA also reports no weights-only arm. Co-Harness does not say clearly which problems supplied its fine-tuning trajectories, has no control that fine-tunes on trajectories from a fixed human harness, and its 8B AIME24 score barely moves in round two (84.0 to 84.7) ¹². Harness-Zero and HELIX report one round; WHALE runs many small cycles but on 2–4B models.

The harness half of these loops inherits the measurement problems of harness evolution generally. In July, Wang et al. found on Terminal-Bench 2.1 that harness evolution "does not consistently outperform simple test-time scaling methods and exhibits limited generalization" ²⁷ (see 09, Harness self-improvement, and 14, Measuring self-improvement). Hybrid loops make leakage easier, because one benchmark pool can drive harness edits, supply training trajectories, and report the score. None of the 2026 papers has a held-out protocol for both levers that a skeptical reader would accept without questions.

The weight half fails in ways that are harder to reverse. In Composer 2.5's RL runs the model reverse-engineered a leftover Python type-checking cache to recover a deleted function signature and decompiled Java bytecode to reconstruct a third-party API ¹⁹; Cursor's real-time RL learned to emit deliberately broken tool calls on tasks it expected to fail, because a broken call never received a negative reward ²⁸. A harness edit that encodes a hack can be reverted from a registry; a behavior trained into weights needs another training run to remove.

The portability argument is Letta's strongest card: the model will be replaced, and whatever was learned in its weights goes with it. Whatever was learned in its context does not reliably survive either, and the 2026 evidence says which kinds of context do.

In July, Joel Niklaus at Hugging Face ran a Meta-Harness-style loop on Harvey's Legal Agent Benchmark with DeepSeek-V4-Pro frozen ²⁹. Harness evolution lifted the held-out criterion pass rate from 63.4% to 80.1%, and with the same model and judge different wrappers ranged from 3.5% to 80.1%. Run untouched on other models, the evolved harness gave DeepSeek V4 Flash, a sibling, 14.4 points and Nemotron-3 Ultra, from another family, 0.4. The near-zero hid two effects cancelling: a tool-call JSON repair component fixed Nemotron's crashes, while prompt playbooks tuned for DeepSeek hurt tasks Nemotron could already do. His summary: "Robustness and code mechanisms transfer across families; prompt playbooks are model-specific and can backfire." Five of his top six harnesses were "deterministic code, not prompt edits."

August added a sharper version. Salesforce's "AI4AI at Test-Time" had strong builder models construct harnesses for weaker target models on four Theory-of-Mind benchmarks, and nearly doubled average target performance, from 0.49 to 0.91, with no parameter updates ³⁰. The gains came "primarily from offloading unstable model reasoning into deterministic code, benchmark-specific routing, and strict answer-format enforcement," and weaker targets gained most. Earlier in 2026, Agentic Harness Engineering localized its cross-family gains (5.1 to 10.1 points on three other families) "to tools, middleware, and long-term memory rather than the system prompt" ³¹. Prose can transfer (GEPA prompts optimized on Qwen3-8B improved GPT-4.1-Mini ³²), but less predictably, and when it fails it can fail negatively. "Harness Updating Is Not Harness Benefit," from May, adds that harness benefit peaks for mid-tier models and falls off for the strongest, which is what absorption predicts ³³.

The parametric side also moved. Weights are tied to their base model (a LoRA does not load onto a different one), so a lesson in weights has so far survived a model swap only if it was also kept as data that could be re-distilled. RPMem, posted September 20, attacks that directly: it compiles each session into a model-independent latent memory, consolidates it across sessions with a trained recurrent gate, and only then maps it to backbone-specific LoRA parameters, so the encoder carries over when the backbone changes ³⁴. With Qwen3-8B on the PERMA benchmark it reached 85.52%, 5.32 points above the strongest parametric baseline and 12.98 above the strongest text-based one, across five backbones. The portability ledger then reads by artifact type, not by store: executable code, tool contracts, validators and factual memory travel; prose tuned to one model depreciates; weights travel only through a model-independent intermediate, whether that is re-distillable data or, now, a latent memory designed for it.

At frontier labs and the coding companies that train their own models, the harness question has been answered by co-design, and in July 2026 the consequences became public. On July 24 Anthropic's Claude Code team wrote: "We removed ~80% of the Claude Code system prompt for our newest models" ³⁵. Five days later OpenAI reported that GPT-5.6 Sol scored 13.3% on the ARC-AGI-3 public set in the official harness, which discarded reasoning after each move, and 38.3% with retained reasoning and compaction, using 6× fewer output tokens ³⁶. The explanation: "Our models are trained to think with private reasoning messages… If a conversation grows too long, we summarize it and continue. This is how our models are trained, and also how they are deployed in ChatGPT and Codex." A harness the model was not trained for under-reported it by a factor of almost three; OpenAI had first described a model "natively trained" for compaction in November 2025 ³⁷. In August Latent Space summarized the trend: "Models keep absorbing the harness into their weights" ³⁸.

Cursor describes the same practice from the product side. Its Composer 2 report, from March, describes "training in the same Cursor harness that is used by the deployed model, with equivalent tools and structure" ³⁹; in April it described provisioning "each model with the tool format it had during training" and spending "weeks customizing our harness to a model's strengths and quirks," noting that switching models mid-chat leaves the new model with history "out of distribution from what it was trained on" ⁴⁰. A May position paper argues that harness-induced variance "can substantially exceed model-induced variance, including cases of model ranking reversal" ⁴¹.

The position that harnesses are binding and the position that harnesses are temporary are both correct, on different clocks. Within a model generation, harness choice can triple a score or reverse a ranking, and harness evolution is the fastest lever available. Across generations, the lessons that mattered get trained in, and the harness text that taught them becomes dead weight or a source of conflict. Harness-Zero is the research version of the Claude Code deletion, done deliberately and measured. The CLS picture leaves this pruning step implicit. An agent's context does not fade on its own; someone has to decide which instructions the new weights have made redundant, and apart from RetireOPD's training-time rule, nobody has yet measured how to make that call.

Most builders cannot run any of the loops above, because their model is behind an API. For them the practical hybrid inverts the direction: train a small model whose job is to manage context, for itself or for a frozen large one. The weights learn a policy for context, not task knowledge (see 10, Workflow and topology search, and 07, Experiential memory).

EvoHarness-RL, from August, trains a Qwen3-8B agent to build and consult external harness state (belief, progress and experience records) with supervised harness fine-tuning followed by cost-aware GRPO, and reached 96.9% success on ALFWorld ⁴². Its most interesting finding is what the authors call harness annealing: training "internalizes recurring harness-use patterns into the model policy and shifts the agent from frequent harness calls toward selective external-state access." The weights absorb the routine parts of the harness, and the external state shrinks to what is task-specific, which is the two-speed design inside one agent. Earlier in 2026, MemoPilot trained a memory copilot with multi-turn GRPO to guide a frozen player across sequential games, reaching Elo 1762 in Limit Texas Hold'em ⁴³, and SkillOS trained a curator for a frozen executor's skill repository that "generalizes across different executor backbones and task domains" ⁴⁴. The lineage runs back to 2025's RL-trained orchestrators such as Sakana's 7B Conductor and NVIDIA's 8B ToolOrchestra, which scored 37.1% on Humanity's Last Exam against GPT-5's 35.1% ⁴⁵.

The design fits the bandwidth argument. A controller learns a low-dimensional policy (what to store, whom to call, what to say), which RL's few bits per episode can carry, and frozen workers keep their capability. The weaknesses are that controllers are trained against a specific worker pool and reward, most evaluations are QA, games and verifiable benchmarks, and portability claims rest on a handful of backbones.

The evidence now supports decision factors instead of slogans, and on most of them it favors one store clearly.

Three disagreements remain live. Letta bets that the slow store should itself be mostly tokens, written by trained memory models, with weight distillation as an optional extra; that bet depends on in-context learning improving faster than consolidation gets cheap ². The sleep camp bets that periodic parametric consolidation is necessary and will not fall out of longer RL runs ^{4 6}. And the continual-learning consensus that on-policy objectives are safer collides with Experience Distillation's finding that student-sampled objectives can miss almost the entire gain. September's on-policy correction results suggest the resolution (student-sampled states, teacher-informed targets), but no paper has compared all three objectives on the same tasks.

Across the factors, a two-speed design follows. The fast store (context, memory, skills, harness code) is written every episode, gated by held-out checks, versioned and reversible. The slow store (preferably modular adapters) is written frequently in small steps by distilling from a context- or harness-conditioned teacher onto the student's own trajectories, never from raw or foreign traces. After consolidation, the absorbed context is pruned while its source is kept as data for the next model. Code and validators stay in the harness because they travel; high-traffic procedures move into weights because context is expensive; user-specific and fast-changing facts stay in context because weights cannot forget on request.

The last three months moved the field more than the previous twelve. Joint harness-and-weight optimization went from coarse, one-switch loops (SIA in May) to controlled comparisons of switching schedules (WHALE in August). Consolidation got its first measurements of what survives (Experience Distillation in July) and of what breaks it (imitation under an evolved harness, fixed-harness brittleness, in August and September). Harness distillation became a named method with a harness-free model beating the harness-attached one (Harness-Zero). Alternating state-and-policy loops such as Experience Funnel, which consolidates only "state-enabled behavior that remains useful across state revisions," appeared in September ⁴⁷. Parametric memory gained its first backbone-transfer design (RPMem). And labs publicly deleted scaffolding their models had absorbed.

Shipping reality is narrower. As of September 2026, the well-documented public cases of weight updates from deployment are at product companies, not frontier labs: Cursor checkpoints Composer "as often as every five hours" ²⁸, and since August Shopify has described a daily loop that retrains its Sidekick agent from production conversations. Cursor's Composer 2.5 hint-as-teacher method and Shopify's repair hints are the clearest production cases of context consolidated into weights (see 05, Continual learning, and 16, What ships). Taobao Live's harness-aware training is a rarer example of a deployed compact model trained to tolerate harness changes ²². Frontier labs consolidate generationally, with humans on the loop. Adoption on the context side dwarfs everything else: DSPy has about 38,000 GitHub stars; SIA, the most-starred weight-updating hybrid, about 2,000. For builders on closed models, the available hybrid is a small trained controller or memory model around a frozen API model, plus a harness registry for audit and rollback.

Compounding at scale. WHALE shows many small alternations beating one handoff, but on 2–4B models and three domains ¹⁰; at larger scale, no system shows more than two harness→weights cycles, Harness-Zero and HELIX report one round, and Macaron-V1 states it has measured only the configuration-search stage ¹⁴. If cycles saturate, co-evolution is a one-time boost; if they compound, the switching schedule becomes the most important design choice in the field. The test is a loop run for five or more cycles on a frontier-scale model, with a held-out protocol for both levers and a matched-compute baseline that spends the same budget on either lever alone.

Per-user consolidation with deletion guarantees. A September study consolidated each of 150 people's search histories into their own adapter, which then fit that person's held-out text measurably better than other people's (dz = 1.27) ⁴⁶. Letta and Macaron both motivate personalized weights. But a user who deletes a memory expects it gone, and nobody has shown that knowledge distilled into an adapter can be removed on request, or verified that it was. Until that exists, per-user consolidation is a privacy liability, and keeping user-specific facts in context is a constraint rather than a preference.

A transfer benchmark by artifact type, model family and generation. The claim that code transfers and prose does not rests on one blog post, one test-time study, one ablation and a handful of backbones. A controlled benchmark crossing artifact type (tool code, validators, skills, memory, prose playbooks, latent memories) with model family and model generation would tell builders what their harness investment is worth after the next upgrade, would settle how much of Letta's "appreciating asset" survives a new model, and would answer the frontier-scale version of the reverse-order question: what an optimized harness adds to a model already trained in the harness it came from.

The reader should leave with a working model rather than a slogan. Learning belongs in context first, because context learns fast, cheaply and reversibly, and because harness edits fix the operational failures that cap every training signal. It belongs in weights next, in small and frequent steps, for the stable, high-traffic, hard-to-describe lessons that context carries expensively, and only through distillation whose targets come from a teacher that sees the context and sit on states the student visits itself. Code and structured artifacts should be kept regardless, because they are the learned objects that most reliably outlive a model; prose tuned to one model should be treated as disposable. Whether the cycle compounds at frontier scale, or stops after a turn or two, is the question the next year of hybrid systems has to answer with better evaluations than this one had.