waggle-os/docs/paper/2026-07-09-tri-benchmark-outline.md

Tri-Benchmark Memory-Systems Paper — OUTLINE (for team-lead approval)

Status: outline only, no section prose. Awaiting KORRO human-in-the-loop sign-off before any section is written. Deliverables covered: (1) 3 candidate titles, (2) thesis + worldview, (3) full section outline with per-section topic sentences + figure/table + evidence source, (4) figure/table plan, (5) claim-evidence map, (6) blog-post variant. Primary evidence base: benchmarks/results/MEMORY-BENCHMARKS-CONSOLIDATED-2026-07.md; docs/paper/2026-06-12-locomo-sota-arxiv-draft.tex; benchmarks/results/beam/{mem0-0641-pipeline-analysis,forensics-p2p-losing-abilities,forensics-retrieval-headroom,forensics-temporal}.md; benchmarks/results/locomo-sota-2026-06/; LongMemEval RESULTS.md (hive-mind workspace).

---

1. Candidate Titles

Per title.md: write one of each type; the best is often the one first rejected. Each names the three benchmarks and/or the striking number, avoids "Novel/Towards/On the."

1. (Provocation) "Clever Memory Loses: A Single Simple Substrate Is State of the Art on LoCoMo, LongMemEval, and BEAM"

   • Rationale: leads with the counter-intuitive claim (structure/distillation/graphs lose), then earns it with three named incumbents. The provocation is the entire thesis in five words.

2. (Concept name) "Conflict-Aware Raw-Turn Memory: State-of-the-Art Long-Term Recall Without Distillation, Graphs, or Routing"

   • Rationale: installs the concept we want cited — conflict-aware raw-turn memory — and the subtitle names the three "clever" mechanisms it beats. The name is the linguistic currency (impact.md Principle 5, ideation.md Protocol 7).

3. (Descriptive precision) "Simple Memory Beats Clever Memory: Three Long-Term-Memory SOTAs and 37 Falsified Interventions Under Each Incumbent's Own Protocol"

   • Rationale: the understated-confidence register (title.md Type 3) that carries the two stickiest numbers — three SOTAs, 37 falsified levers — and the fairness hook (own protocol) that pre-empts the "cross-lab noise" reviewer.

(Working shortlist favorite: #1 as title, #2's concept name threaded through the paper as the named contribution. Decision deferred to team lead.)

---

2. Thesis and Worldview (per impact.md + ideation.md)

One-sentence thesis (no jargon):

Across all three major long-term-memory benchmarks, a deliberately dumb memory — the original dated conversation turns, retrieved per conversation, with contradictions kept rather than reconciled — beats every state-of-the-art system that distills, structures, graphs, or routes its memories, and we measured 37 "clever" additions, ~90% of which made it worse.

Worldview destroyed: "Better agent memory means cleverer memory — richer structure (knowledge graphs, entity reconciliation, atomic-fact distillation), learned routing, and test-time ensembling. The frontier is engineering more intelligence into the memory path."

Worldview installed: "The memory path should stay dumb. Distillation, reconciliation, and routing are mostly self-inflicted information loss; a substrate that preserves raw dated turns and surfaces conflicts instead of silently resolving them is the state of the art. This is the Bitter Lesson for agent memory."

Memorability hooks (deploy ≥2, impact.md P5):

• The Number: "37 interventions measured, ~90% falsified." Secondary: "+23pp on the benchmark's hardest ability (contradiction)."
• The Analogy: Sutton's Bitter Lesson — hand-engineered structure loses to simple methods that leverage the raw data/compute. Here: distilled facts / KGs lose to raw dated turns.
• The Name: conflict-aware raw-turn memory (contribution) and the falsification ledger (the negative-results artifact).

Inevitability arc (impact.md P4) — section-to-phase mapping is enforced in §3 below.

---

3. Full Section Outline (arXiv preprint)

Convention below: each section lists 3-6 topic sentences (first-sentence-of-paragraph messages, per introduction.md checklist), the figure/table that lands there, and the evidence source file/number.

3.1 Abstract (Template B: Challenge → Insight → Contribution, abstract.md)

• Topic 1 (task): Long-term conversational memory — answering over weeks of prior dialogue — is the load-bearing capability for durable AI assistants, and three benchmarks (LoCoMo, LongMemEval, BEAM) are the field's rulers.
• Topic 2 (challenge): Every published leader adds structure to the memory path — atomic-fact distillation, temporal knowledge graphs, write-time entity reconciliation, learned routing, test-time ensembling.
• Topic 3 (insight): These are lossy. Distillation strips the dates and specifics the rubrics reward; write-time reconciliation silently resolves contradictions the rubrics want surfaced.
• Topic 4 (contribution): One simple substrate — per-conversation minds, verbatim dated raw-turn retrieval, and a conflict-preserving answer policy — is SOTA on all three under each incumbent's own published protocol (LoCoMo 86.49 vs 81.95; LongMemEval 95.01 macro vs 94.87; BEAM 0.6482 / 74.0% vs 0.6409 / 70.1%).
• Topic 5 (benefit / negative results): We measured 37 constructive interventions across two programs; ~90% lost to the simple substrate — negative results reported as a first-class contribution (the falsification ledger).
• Topic 6 (experiment summary): The single differentiator is conflict-awareness: +23pp on BEAM contradiction_resolution. All protocols, per-question artifacts, and the empty-answer-heal disclosure are released.
• Figure/Table: none (abstract). Evidence: consolidated lines 7-11, 66-89.

3.2 Introduction (logic map from introduction.md; Version A4 "open with challenge")

• Topic 1: A system that cannot recall what its user said three weeks ago cannot be a durable collaborator; LoCoMo, LongMemEval, and BEAM operationalize this and now anchor a public leaderboard race.
• Topic 2 (SOTA + root issue): The leaders differ in representation (Mem0 atomic facts, Zep temporal KG, Memori triples, Mastra, mem0-platform reconciled facts) but agree on one premise: cleverer structure on the memory path is the path forward.
• Topic 3 (the crack in the worldview): Yet each incumbent's own weakest ability traces to its own cleverness — mem0's write-time reconciler is its worst on contradiction (0.357); distillation is worst on fine-grained recall.
• Topic 4 (our contribution, impact.md "make the reader the hero"): We show a single dumb substrate is SOTA on all three, and — more useful to practitioners — we publish 37 measured interventions so the community stops paying the distillation/routing tax.
• Topic 5 (contribution list): (i) three same-protocol SOTAs; (ii) conflict-aware raw-turn memory as the named mechanism; (iii) the falsification ledger; (iv) an evaluation-protocol-fidelity methodology that exposed leaderboard-metric chaos.
• Topic 6 (why now / hindsight, ideation.md Test B/C): The enabling shift is long-context answerers (gpt-5) that can consume raw turns directly — making distillation not just lossy but unnecessary; in hindsight, "of course you shouldn't throw away the data before the model reads it."
• Figure/Table: Table 1 (headline tri-benchmark summary). Evidence: consolidated lines 7-11; tex Introduction §.

3.3 Related Work

• Topic 1 (memory systems on LoCoMo): Mem0 (atomic facts), Zep/Graphiti (temporal KG), LangMem (profiles), Memori (timestamped triples, prior SOTA 81.95); MemR3 is an agentic multi-round retrieval pipeline, classed separately. (reuse tex Related Work verbatim-adjacent.)
• Topic 2 (LongMemEval line): Mastra (94.87 macro) and mem0-platform (93.4, GPT-5 judge) as incumbents; note the judge-model divergence up front.
• Topic 3 (BEAM line): mem0 memory-platform pipeline (2-turn-chunk write-time reconciliation, top-200 dated facts, chronological render) as the ICLR-2026 reference (0.6409 / 70.1%).
• Topic 4 (the common commitment we break): Every system above commits to a dominant derived representation and to write-time resolution of conflicts; none retains raw dated turns as the primary answer substrate or preserves both sides of a contradiction.
• Topic 5 (where we differ): We layer representations over a common frame store but let the answer path draw verbatim raw turns first; the benchmark categories are won by different representations, and a substrate that serves all simultaneously beats any single-representation system. (tex Related Work "Where Hive Mind differs.")
• Figure/Table: none. Evidence: tex Related Work; consolidated line 87; mem0-0641 pipeline analysis §"The pipeline, step by step."

3.4 The Substrate (Architecture) (impact.md arc Phase 4 "resolution")

• Topic 1 (storage core): Memory is stored as typed frames (Information / Preference / Belief) plus verbatim per-turn frames in per-conversation "minds"; retrieval is hybrid dense (nomic-embed) + FTS5 BM25 with RRF fusion and optional cross-encoder rerank — fully local.
• Topic 2 (the primary lane is raw): Unlike the incumbents, the answer context is built from dated verbatim raw turns retrieved per conversation, date-stamped [YYYY-MM-DD]; distilled layers are auxiliary, not the substrate.
• Topic 3 (write-time dating, the one thing we DO transform): Relative dates are resolved at write time ("yesterday" in an 8 May session → 7 May event) — the single distillation that survived, because it adds information rather than removing it.
• Topic 4 (conflict-preserving policy — the differentiator): The answer policy retains both sides of a conflict and instructs the model to state the contradiction and ask which is correct, rather than reconciling at write time; this is the mechanism behind the BEAM contradiction win.
• Topic 5 (answer policy): Conditional abstention (commit on factual spans when an anchor exists; abstain only when truly absent), granularity-calibrated dates, commit-to-one-option — tuned to what rubrics reward, not to prose quality.
• Topic 6 (one substrate, three protocols): The same substrate is evaluated under three different answerers/judges by swapping only the protocol harness — no benchmark-specific memory engineering.
• Figure/Table: Figure 1 (substrate architecture: minds → I/P/B frames + raw dated turns → hybrid retrieval → assembler lanes → conflict-aware answer policy). Evidence: tex §"The Hive Mind substrate"; consolidated lines 59-64.

3.5 Evaluation Protocol Fidelity (the metric-audit methodology) (impact.md arc Phase 2-3 "tension + diagnosis"; the honest-metrics thread)

• Topic 1 (why a whole section): Cross-lab long-term-memory numbers are noise; the only defensible comparison is same judge, same protocol, reproduced pipeline — so we first reproduce each incumbent before comparing.
• Topic 2 (LoCoMo fidelity): We reproduced Memori's own pipeline to 81.98 vs their published 81.95 before any comparison, restoring image-caption input parity (protocol compliance, not architecture).
• Topic 3 (LongMemEval metric chaos): The leaderboard mixes metrics — Mastra's 94.87 is macro (category mean), mem0's 93.4 used a GPT-5 judge not the official GPT-4o; we report both macro (95.01) and micro (93.60) under the official GPT-4o judge, the most conservative cell on the board.
• Topic 4 (BEAM metric chaos): mem0's shipped 0.641 result file is gpt-5/gpt-5, though the repo README defaults imply gpt-4o/gpt-4o — an answerer/judge confound we neutralize by matching gpt-5/gpt-5, and we report both pass-rate (headline) and avg-score (within noise).
• Topic 5 (full disclosure principle): We disclose the empty-answer-heal bug (56 rows, reasoning-budget cap), keep the pre-heal snapshot, and report judge-noise bands (SE ≈ 0.014 on BEAM avg) — the anti-cherry-pick posture.
• Figure/Table: Table 6 (protocol-fidelity / metric-audit matrix: per benchmark — what the leaderboard reported, judge model, macro-vs-micro, README-vs-actual, our reported cell). Evidence: mem0-0641 pipeline analysis §"Headline finding"; consolidated lines 38-40, 91-96; forensics-p2p §2 (empty-answer bug).

3.6 Results (three benchmarks) (impact.md arc Phase 4)

• Topic 1 (LoCoMo): 86.49% (1332/1540) under Memori's protocol, +4.54pp (one-sample z=4.64, p<10⁻⁵), leading every category among memory systems and statistically at the full-context ceiling on open-domain; verified drift-free (fresh regen on HEAD reproduced 86.49 exactly).
• Topic 2 (LoCoMo field correction): We correct a column-scrambling error propagating through the literature (LangMem's cited "temporal 86.92" is its single-hop; true temporal 61.06) and re-run Mem0 same-judge (we lead every category, temporal +30.8pp).
• Topic 3 (LongMemEval): 95.01 macro / 93.60 micro (468/500), SOTA under the leader's own macro metric; the arc from 75.9 was won by four answer-side survivors, not by richer memory.
• Topic 4 (BEAM headline): 0.6482 avg / 74.0% pass (700 Q, gpt-5/gpt-5, top-30 raw turns) vs mem0 0.6409 / 70.1%; pass-rate +3.9pp (~2.3 SE, statistically real), avg +0.007 within noise (both reported).
• Topic 5 (BEAM differentiator): contradiction_resolution 0.588 (87.1% pass) vs mem0 0.357 (48.6%) — +23pp on the benchmark's hardest ability — plus abstention +6.8pp and information_extraction +4.0pp; the losses are concentrated in coverage abilities (summarization −7.4, temporal −6.1).
• Topic 6 (efficiency framing): BEAM SOTA at top-30 raw turns beats mem0's top-200 facts (4× depth buys them only +0.037) — compactness of raw beats depth of distilled; total program spend ≈ $160.
• Figure/Table: Table 2 (corrected LoCoMo landscape), Table 3 (LongMemEval intervention arc), Table 4 (BEAM per-ability vs mem0). Evidence: consolidated §1-§3 (lines 17-89); tex Table tab:field; mem0-0641 §"Second finding" (top-200 vs top-50).

3.7 The Falsification Ledger (negative results) (the first-class contribution; ideation.md Protocol 3 territory map)

• Topic 1 (framing): We treat negative results as the product: 37 constructive interventions across two independent programs (23 LongMemEval + 14 BEAM), ~90% falsified against the simple substrate.
• Topic 2 (compression/distillation loses): Distilled-fact retrieval (k=100 → 0.354, k=200 → 0.382) and every raw+fact hybrid (0.498-0.540) lost — facts are net-harmful at any mix because they strip dates/numbers; the session-outline preamble diluted summarization 0.38→0.16.
• Topic 3 (routing loses): Gold-blind ability routing hit a 40% classifier ceiling (0.584 < 0.600) — abilities are not text-identifiable; coverage-shaped/stratified retrieval lost to similarity-dense clusters (summ 0.38→0.24).
• Topic 4 (prompt engineering plateaus): Strict-abstention (0.542), exhaustive-enumeration (−0.048; summ −0.119), and temporal-commit (+0.023, below re-roll variance) prompts all failed to beat the plain conflict-aware v2 prompt.
• Topic 5 (ensembling loses): Answer-merge self-ensemble (0.668/0.674) was within noise for 2× the cost; on LongMemEval all four surviving levers were answer-side (voting vh5, KG split-vote guard klc, preference critic pfc, routing-matrix completion h3u) — never memory-structure levers.
• Topic 6 (the convergent conclusion): Two programs, different benchmarks, same verdict — the local optimum is plain dated raw-turn retrieval + a conflict-aware prompt; cleverness on the memory path is measured self-harm.
• Figure/Table: Table 5 (the falsification ledger — 37 rows: lever, program, score, why it lost, verdict), Figure 2 (falsification scatter: intervention complexity vs Δscore, ~90% below the raw-turn baseline). Evidence: consolidated BEAM 14-lever table (lines 102-117), lines 44-45 + 124-126; LongMemEval RESULTS.md.

3.8 Analysis (why simple wins) (impact.md arc Phase 5 "map the new territory")

• Topic 1 (why simple wins — the mechanism): Retrieval headroom forensics show 88-93% of failed nuggets were already retrieved at rank ≤30 — the bottleneck is answer-side synthesis, not retrieval depth, so adding structure/depth cannot help and distillation only removes what the answerer needed.
• Topic 2 (contradiction case study — the differentiator): mem0 solves knowledge_update via write-time UPDATE/DELETE but that same reconciler makes contradiction its floor (0.357); our retain-both-sides substrate scores 0.588 — the trade is intrinsic to write-time resolution, not a tuning gap.
• Topic 3 (ability-level tradeoffs — honest): Where we lose (summarization −7.4, event_ordering −4.2), the forensic cause is coverage/clause density, not memory quality — mem0's enumerative fact dumps (1067 vs 600 words, 48 vs 29 bullets) mechanically harvest more compound-nugget clauses; largely an answer-verbosity artifact of the rubric.
• Topic 4 (temporal is disposition, not knowledge): The BEAM temporal gap (−6.1) is 94% duration arithmetic; 74% of the loss is over-abstention + wrong-anchor (single-date-selection), and a chunk is irreducible gpt-5-self-judge phrasing variance — not a memory-representation deficit.
• Topic 5 (what this means for the field): The load-bearing levers are (a) keep the raw data, (b) resolve dates at write time, (c) preserve conflicts — everything else the field is building is optional at best, harmful at worst.
• Figure/Table: Figure 3 (contradiction mechanism: write-time reconcile collapses to one side vs retain-both-surface; paired with the 0.588 vs 0.357 bar), Table 7-inline (retrieval-headroom rank distribution, small). Evidence: forensics-retrieval-headroom (rank table); forensics-p2p §5-§8 (coverage/verbosity); forensics-temporal §2-§3; mem0-0641 §"Ingestion."

3.9 Limitations (ideation.md Protocol 6 anti-paper pre-emption)

• Topic 1 (summarization/coverage gap): We are genuinely behind on coverage-shaped abilities; whether the raw substrate can close it without the enumeration tax (which we showed backfires) is open.
• Topic 2 (judge noise): All scores are LLM-as-judge; we mitigate with fixed judges, archived per-question judgments, reproduced baselines, and reported noise bands — but absolute numbers are protocol-relative, and some BEAM deltas sit inside gpt-5-self-judge variance.
• Topic 3 (single-substrate / single-answerer scope): Each benchmark uses one answerer per its protocol; model-generality of the raw-turn design is untested, and LoCoMo token cost (3747 vs 1294) means there is no free ≤1500-token operating point — the raw lanes are load-bearing.
• Topic 4 (token/cost honesty): Raw-turn top-k is cheap at k=30 but scales super-linearly (k=100 ≈ 88K ctx tokens/question); distillation's appeal is cost, not quality — we win quality, they win the efficiency axis on LoCoMo.
• Topic 5 (the empty-answer bug, disclosed): The healed 56-row reasoning-budget bug is disclosed with a pre-heal snapshot; the residual post-heal gaps are the honest comparison.
• Figure/Table: none. Evidence: tex Limitations §; forensics-retrieval-headroom (k-cost table); forensics-p2p §2; consolidated lines 91-96.

3.10 Conclusion (impact.md P8 call-to-action; ideation.md Protocol 3)

• Topic 1 (restate the one thing): One dumb substrate — dated raw turns, per-conversation, conflicts preserved — is state of the art on all three long-term-memory benchmarks.
• Topic 2 (the bitter lesson): 37 measured interventions say the field is over-engineering the memory path; the wins are in not discarding data and not resolving conflicts early.
• Topic 3 (call to action): All protocols, per-question artifacts, the falsification ledger, and the harness are open; the community can stop re-deriving the distillation/routing tax and audit its own leaderboard metrics.
• Topic 4 (open territory): (i) Can conflict-preservation be made cheap enough to also win coverage abilities? (ii) Is there a write-time transform other than date resolution that adds rather than removes? (iii) Does the raw-turn result hold under weaker/cheaper answerers, or is long-context the true enabler? (iv) What is the right rubric for "surfaced a contradiction correctly" vs "committed and got lucky"?
• Figure/Table: none. Evidence: consolidated line 124-126; tex Conclusion §.

---

4. Figure / Table Plan (~6 tables, ~3 figures)

#	Artifact	Lands in	Exact data source
Table 1	Headline tri-benchmark summary (ours / best competitor / protocol / verdict)	Intro	consolidated lines 7-11
Table 2	Corrected LoCoMo landscape (per-category memory systems + MemR3 + full-context ceiling)	Results §3.6	tex tab:field (lines 389-403)
Table 3	LongMemEval intervention arc (75.9 → 95.01 macro / 93.60 micro, lever labels vh5/klc/pfc/h3u)	Results §3.6	consolidated lines 41-43; RESULTS.md
Table 4	BEAM per-ability vs mem0 (10 abilities + overall avg + pass rate)	Results §3.6	consolidated lines 68-81
Table 5	The Falsification Ledger (37 levers: lever / program / score / why-lost / verdict)	§3.7	consolidated BEAM table lines 102-117 (14) + RESULTS.md (23)
Table 6	Protocol-fidelity / metric-audit matrix (leaderboard-claim / judge model / macro-vs-micro / README-vs-actual / our cell)	§3.5	mem0-0641 §"Headline finding"; consolidated lines 38-40
Figure 1	Substrate architecture (minds → I/P/B + raw dated turns → hybrid retrieval → assembler lanes → conflict-aware policy)	§3.4	tex §"The Hive Mind substrate"; consolidated 59-64
Figure 2	Falsification scatter (intervention complexity vs Δscore; ~90% below the raw-turn baseline line)	§3.7	consolidated 102-117; RESULTS.md
Figure 3	Contradiction mechanism diagram + bar (write-time reconcile → one side vs retain-both-surface; 0.588 vs 0.357)	§3.8	mem0-0641 §"Ingestion"; consolidated line 70, 86-89
(inline mini-table)	Retrieval-headroom rank distribution (≤30 answer-side 88-93%)	§3.8	forensics-retrieval-headroom rank table (lines 40-44)

---

5. Claim-Evidence Map (every headline claim → file/number)

#	Headline claim	Number	Evidence source
C1	LoCoMo SOTA	86.49% (1332/1540), +4.54pp, z=4.64, p<10⁻⁵	consolidated line 9; tex abstract + tab:field
C2	LoCoMo reproduced baseline before comparing	81.98 vs published 81.95	tex §Experimental setup / Introduction contrib 1
C3	LoCoMo drift-free	fresh HEAD regen reproduced 86.49 exactly	consolidated line 9
C4	LoCoMo field-table correction	LangMem "temporal 86.92" = single-hop; true 61.06	tex §"corrected LoCoMo field"
C5	Mem0 re-run same-judge	lead every category, temporal +30.8pp	consolidated line ~; tex tab:field footnote (+32.7pp)
C6	LongMemEval SOTA (macro)	95.01 macro vs Mastra 94.87	consolidated line 10; lines 38-43
C7	LongMemEval micro reported too	93.60 micro (468/500), GPT-4o judge	consolidated line 10, 43
C8	LongMemEval levers	23 measured, 4 adopted (vh5,klc,pfc,h3u), 19 falsified	consolidated lines 44-45; RESULTS.md
C9	BEAM SOTA (pass rate)	74.0% (518/700) vs 70.1% (491/700), +3.9pp ≈2.3 SE	consolidated lines 11, 81-84
C10	BEAM avg reported honestly	0.6482 vs 0.6409, +0.007 within noise SE≈0.014	consolidated lines 80, 83-84
C11	Contradiction differentiator	contradiction_resolution 0.588 vs 0.357 = +23pp	consolidated line 70, 86-89
C12	Why: mem0 reconciler is its own floor	write-time UPDATE/DELETE silently resolves; their weakest 0.357	mem0-0641 §Ingestion (lines 43-49, 78)
C13	Depth is not the lever	mem0 top-200 vs top-50: 4× depth = +0.037	mem0-0641 §"Second finding" (lines 28-37)
C14	Simple wins is answer-side	88/90/93% of failed nuggets already at rank ≤30	forensics-retrieval-headroom (lines 40-63)
C15	Our losses are coverage/verbosity	mem0 answers 1067 vs 600 words, 48 vs 29 bullets	forensics-p2p §5 (lines 82-101, 263)
C16	Temporal gap = disposition not knowledge	94% duration arithmetic; 74% over-abstention + wrong-anchor	forensics-temporal §2-§3 (lines 72-118)
C17	37 levers, ~90% falsified	14 BEAM (table) + 23 LME	consolidated 102-117; 44-45; 124-126
C18	Distillation is lossy	distilled k=100 → 0.354 vs 0.448 raw	consolidated line 104
C19	Empty-answer bug disclosed	56 rows healed; pre-heal snapshot kept	consolidated 91-96; forensics-p2p §2
C20	Metric audit: hidden judge swap	mem0 BEAM 0.641 file = gpt-5/gpt-5 (README implies gpt-4o)	mem0-0641 §"Headline finding" (lines 5-24)
C21	Cost	total BEAM program ≈ $160; LoCoMo 3747 vs 1294 tok/q	consolidated line 13; tex Limitations

Open dependency to resolve before writing: C8 (LongMemEval 19 falsified levers) and Table 3/5 LME rows are sourced from RESULTS.md in the hive-mind workspace, not the waggle-os filesystem — pull that file into benchmarks/results/longmemeval/ for citable line numbers before drafting §3.6-§3.7.

---

6. Blog-Post Variant Outline (800-1500 words)

One insight: "Simple memory beats clever memory — we measured 37 ways." Title candidates: "We Tried 37 Ways to Make AI Memory Smarter. 33 Made It Worse." / "The Bitter Lesson of Agent Memory." Register: title.md blog rule — click-worthy, not clickbait; lead with the number and the counter-intuitive result.

• Hook (≈120 w): The industry is racing to build cleverer memory for AI agents — knowledge graphs, fact distillation, entity reconciliation. We benchmarked a deliberately dumb alternative and it won on all three major benchmarks. Then we tried 37 ways to make it smarter; almost all made it worse.
• The setup (≈150 w): Three benchmarks (LoCoMo, LongMemEval, BEAM), each with a published SOTA. The honest-comparison rule: same judge, same protocol, reproduce the incumbent first. One substrate, no per-benchmark memory engineering.
• The result in one table (≈120 w): 86.49 vs 81.95 / 95.01 vs 94.87 / 74.0% vs 70.1% — with the caveat we're transparent about (avg-score is within noise on BEAM; pass-rate is the real win).
• The counter-intuitive core (≈250 w): The falsification ledger. Distillation strips the dates and specifics the questions ask for. Routing hits a classifier ceiling. Ensembling costs 2× for nothing. The one transform that helped — resolving "yesterday" to a real date at write time — adds information instead of removing it. Bitter Lesson analogy.
• The differentiator (≈250 w): Conflict-aware memory. Competitors reconcile contradictions when they write memory — which silently deletes one side. We keep both and surface them. Result: +23pp on the hardest ability. The vivid example: "you said 10 vector-field problems here, and 15 there — which is right?" vs a system that already threw one number away.
• The honesty section (≈200 w): Where we lose (coverage/summarization), why (answer verbosity gaming compound rubrics, not memory quality), and the metric mess we found (macro vs micro, GPT-5 vs GPT-4o judges, README-vs-actual protocols) — plus the empty-answer bug we disclosed.
• The takeaway / CTA (≈150 w): Stop over-engineering the memory path. Keep the raw data, date it at write time, preserve conflicts. Everything is open — protocols, per-question records, the ledger. Link to the arXiv preprint.
• Sticky elements: the number (37 / ~90%), the analogy (Bitter Lesson), the contradiction example (10 vs 15 problems), one figure (the falsification scatter, Figure 2 reused).

---

7. Pre-Writing Gate Check (ideation.md decision gate — all filled)

1. One-sentence insight: raw dated turns + preserve conflicts beats distill/graph/route — measured on 3 benchmarks, 37 falsified levers.
2. Why nobody before: everyone assumed richer structure = better memory; long-context answerers only recently made raw-turn feeding viable (Why-Now).
3. Hindsight verdict: "of course you shouldn't delete the data before the model reads it, and of course you shouldn't resolve a contradiction the user asked about."
4. Abstraction (verb form, domain-free): preserve the raw signal and the conflicts; transform only when the transform adds information.
5. The name: conflict-aware raw-turn memory / the falsification ledger.

Anti-paper pre-emptions staged (Protocol 6): cross-lab noise → same-protocol reproduction (§3.5); "avg-score is within noise" → we lead on pass-rate and say so (§3.6, C10); "you lose on summarization" → shown to be rubric-verbosity artifact, stated as limitation (§3.8/§3.9); "judge is unreliable" → fixed judge + noise bands + disclosure (§3.5/§3.9).