Automatic memory writes cover different work
This bounded comparison finds automatic writes in all three selected systems, but their supported mechanisms differ: revising problem-solving guidance, extracting conversational facts, and rebuilding access data. The useful contrast is what gets changed and who consumes it afterward.
This is a current-input workshop trial in temporary-bundle mode, not a public landscape publication. Selection is explicitly one generated review each for Dynamic Cheatsheet, Mem0 and Napkin; all other systems and the legacy corpus are excluded. All 3/3 results are code-grounded, with analysis cutoff 2026-09-26. The lens covers memory accumulated or changed through use, its access structures and later consumers. Boundaries differ: Dynamic Cheatsheet’s workflow includes opaque provider state; Mem0 covers synchronous OSS memory with default adapters; Napkin covers the package and optional skills, excluding their host executor.
Snapshot manifest SHA-256: 101754b023abbd2efd352e11c8cc75a234dc81802829c2c2d15199b05f8b3113. Matrix SHA-256: dc82c6f2e5e9084214147782cf7361c64b20e2940139122e5ded38ad5a4d800d. The temporary evidence location is kb/reports/cache/agentic-memory-refresh/2026-09-26-per-value/bundle. Its recorded precommit HEAD, 26fbe60caa5431b958a7f01569b404ffa1978bc9, alone does not retain all input bytes. Publication requires durable retention of the exact inputs.
The full retained results identify every example below; each linked directory names its run ID:
| Result | Reviewed source revision | Result SHA-256 |
|---|---|---|
| AAS-2026-09-26-dynamic-cheatsheet-02 | 5cfe3c37e8e52b1d858d0f3df46e7f17c50991b9 |
69cf228433d975ac518c02f315ac81aaa26a6b29c3e3db63b932e6a199194cfa |
| AAS-2026-09-26-mem0-02 | 94c3fe9f238f3dbf29c9ce98643bd71eb13077cd |
b4a541c5e6a90e37dbb0115bf9e230c41c5fc635023b519a57c447d4b7ffabd1 |
| AAS-2026-09-26-napkin-03 | 7582d6a46f5a11995956e60a59c41a5b242109f1 |
1c4b625211fa033130d7ea2533c214f5e5e791d07707ad96ec8aa34f9e835c53 |
The query ledger counts each system once per query, accepting code-grounded per-value evidence at wired, observed or causally supported strength. Automatic write membership is 3/3 selected systems. Trace-derived writing and push each have positive witnesses in 2/3: Dynamic Cheatsheet and Mem0. Dynamic Cheatsheet has partial coverage on these axes; Mem0 has known coverage. Napkin’s trace extraction is afforded, and its known package read-back profile is pull-only. Its push signal is inapplicable. These dispositions are not observed negatives about excluded hosts. No relevant row is uninspected.
Dynamic Cheatsheet’s supplied curator templates direct replacement of retained guidance from solution trajectories; subsequent generator calls receive that guidance automatically. Template configuration matters: an omitted curator path supplies literal (empty). Its Shared records RTE-2, RTE-3, RTE-6 establish transformation, supply and persistence. CLM-4 observes a wrong answer retained and delivered in later recorded prompts; the generating revision and later model dependence remain unknown.
Mem0’s RTE-1, RTE-5 and BAP-1 instead wire new fact extraction from messages, with prior facts and recent messages automatically supplied to the next extractor. This is push to an internal extraction model; external answering use is only afforded by RTE-3/BAP-2. Ordinary extraction adds facts. Caller-directed replacement and deletion belong to RTE-2. These Shared records prevent equating accumulation with automatic correction.
Napkin’s automatic-write witness is RTE-6: requested access regenerates and persists caches. Its Shared records RTE-3/RTE-4 afford semantic distillation and upkeep through an external skill agent. Wired cache generation cannot upgrade that extraction route. RTE-1/RTE-7 implement requested returns, leaving subsequent agent use outside the package.
Per-value membership also differs from complete-profile evidence. Natural-language and symbolic memory have positive witnesses in 3/3, but only 2/3 representation profiles—Mem0 and Napkin—have known coverage and strong evidence for every member. Dynamic Cheatsheet’s opaque included state prevents completeness. Faithfulness is not-determinable in 3/3; 0/3 qualify as evidenced “no.” Unknown and partial coverage do not mean absence.
Local semantic self-check covered all three full result bodies, canonical examples and limitations. The ledger rerun matched its exact output; bundle/source verification and Markdown validation passed. Performance ranking, causal benefit, complete opaque profiles and prevalence or change across the old corpus are withheld.