Storage automation and agent memory use remain separate findings
This three-system pilot distinguishes automatic storage work from evidence that an agent extracts and consumes retained knowledge. Dynamic Cheatsheet and Mem0 have implemented trace-fed memory writes and automatic supply to later model calls. Napkin has implemented file operations, imports and access-cache writes, while its agent extraction and recall workflows remain afforded by shipped instructions and documented integrations. These are differences in the retained evidence and selected boundaries, not a ranking of the products.
Frozen population
This is a current-input reliability trial captured on 2026-09-27, not a public full-corpus landscape. Selection is exactly the generated Dynamic Cheatsheet, Mem0 and Napkin reviews; the Napkin review was read only after AAS-2026-09-27-napkin-01 completed. All three results are code-grounded, with no doc-grounded rows. A code-grounded result can still contain claimed or afforded values.
| Result and run | Compared boundary | Source revision | Cutoff | Exact result SHA-256 |
|---|---|---|---|---|
| Dynamic Cheatsheet, AAS-2026-09-26-dynamic-cheatsheet-02 | Six workflow variants and caller/benchmark memory, including opaque provider-native state | 5cfe3c37e8e52b1d858d0f3df46e7f17c50991b9 |
2026-09-26 | 69cf228433d975ac518c02f315ac81aaa26a6b29c3e3db63b932e6a199194cfa |
| Mem0, AAS-2026-09-26-mem0-02 | Synchronous OSS Python memory, default adapters and internal context; enclosing agents excluded | 94c3fe9f238f3dbf29c9ce98643bd71eb13077cd |
2026-09-26 | b4a541c5e6a90e37dbb0115bf9e230c41c5fc635023b519a57c447d4b7ffabd1 |
| Napkin, AAS-2026-09-27-napkin-01 | Package, skills, documented consumers and benchmark handoffs; external host execution excluded | 7582d6a46f5a11995956e60a59c41a5b242109f1 |
2026-09-27 | 8484f622335c134237968326c9971d8fbb1e71f97128e411f96b8d46ccba41db |
The manifest hash is 44f4ef62bda8116eeac568daae97207931a6cd4ad93f459151091a00903a76cd; the matrix hash is fb2f5e4554e24e9514afe71ea6c75b7f6e7b12e0f0082de02d92a0d4f097ed06. Trial acceptance records execution and checks. The exact bundle is retained at trial-bundle/ beside this report; findings cite the original retained results above. Use the retained bundle to reconstruct this snapshot; the earlier checkout revision alone does not identify its input bytes.
What the selected evidence supports
Automatic writing has a wired witness in 3/3 selected systems, but implemented trace-fed derived memory has a witness in 2/3, Dynamic Cheatsheet and Mem0. Push to a later model has wired or observed support in those same 2/3. Napkin's trace extraction and push are afforded, so its exclusion from those implementation counts is not a negative finding. The query conjunction establishes co-occurrence within each system; it does not establish causal benefit or that any two arbitrary routes connect. The query ledger retains all 21 exact records, filters, denominators, included runs and exclusions.
Dynamic Cheatsheet's cumulative route sends the current cheatsheet to the generator, then asks a curator to replace it from the problem, output and old sheet. Retained examples show an earlier sheet in the next prompt. That is observed delivery; the same sample contains a wrong solution retained in the sheet. It does not show that recalled content caused a later answer or establish a general error rate. The curator's intended content criticism depends on supplying the inspected templates, and the extractor itself supplies no semantic veto. Evidence: RTE-2, RTE-6, CLM-4, in Shared records.
Mem0's automatic consumer is its own extraction model: each inferred add supplies selected prior messages and fact text before generating additional facts. Public retrieval for an answering application is a separate afforded route. The ordinary extraction path adds facts; explicit caller operations own replacement and deletion. This is why internal push, external pull and correction cannot be inferred from one another. Evidence: RTE-1, RTE-2, RTE-3, RTE-5, BAP-1, BAP-2, in Shared records.
Napkin's automatic imports and compiled search caches account for its wired automatic-write value. Its distill skill separately directs an external agent to turn session findings into lasting notes, preserve reasons and mark new generalizations. Requested overview/search/read is pull; documented whole-note session loading is coarse push, whose host implementation is outside the inspected package. Neither automatic indexing nor a requested search response supplies that missing implementation evidence. Evidence: OBJ-8, RTE-7, RTE-8, RTE-10, RTE-11, in Shared records.
Both natural-language and symbolic retained parts have strong witnesses in 3/3 selected systems. This does not mean their learned outputs are identical: symbolic access metadata and query definitions count in representational form, while distilled form concerns the specific trace-fed transformation. Dynamic Cheatsheet also has partial aggregate coverage because included provider payloads remain opaque. Its supported positives count, but its value sets cannot enter complete-profile comparisons.
Withheld conclusions and checks
The trace_learning axis names a retained trace-fed write route. It establishes neither theory-builder membership nor improved future capacity caused by criticism. All 3/3 results mark recall-dependence testing not-determinable. The candidate query for an evidenced “no” therefore finds 0/3; this cannot be rewritten as “none tested recall” or “none used memory.” No result here supplies a common performance comparison, product ranking, trend, or historical-corpus prevalence estimate. Boundaries and cutoffs differ, and no second comparable snapshot was admitted.
All full bundled results were read in bounded calls without truncation. Query verification reproduced all 21 records and independently checked them against bundled result frontmatter. Local semantic verification traced the examples to their canonical records, source excerpts and limits; no independent agent review occurred. The matrix, table and statistics used the same three explicit review arguments and matched the bundle's input identities. Bundle verification with the externally recorded hash and --source-root . passed before writing; the acceptance record holds the final source recheck and Markdown-validation result.