Memory refresh synthesis trial — 2026-09-26
Historical initial pilot, before ADR 093 adopted per-value evidence. The results, contracts and counts below belong to that frozen first trial.
This workshop trial compares who selects retained context, who changes it, and what checks justify later reliance. The explicitly selected population is Napkin, Mem0 and Dynamic Cheatsheet: three current analyses, all code-grounded (3/3), with cutoff 2026-09-26. Their boundaries differ: Napkin's product includes shipped skills but excludes its external host; Mem0 covers synchronous Python OSS memory; Dynamic Cheatsheet covers local memory and a benchmark controller with external model execution. This is a bounded comparison, not coverage of the legacy corpus.
Napkin leaves the next read with the caller. Its executable route returns a requested folder map, lexical search ranked with backlinks and recency, or full note text. Its distill and tend skills instruct an external agent to extract durable notes, merge overlap and identify contradictions. Those maintenance routes are afforded: code supplies file operations, but the host must invoke the skills and the model must follow them. A requested overview containing project context remains pull; excluded host injection cannot supply a push claim. Persistence and retrieval are wired, while compliant distillation and later behavioral use remain uninspected. See Napkin's shared records, RTE-2, RTE-3, RTE-4, RTE-5 and BAP-1.
Mem0 wires retained context into its own next extraction call. Inferred add selects recent messages by scope identity and old memories by embedding similarity, then supplies their text to the extractor. Both extracted facts and procedural summaries can enter that later context. This is push to the extraction model; requested SDK retrieval affords a separate pull route to an application. The inferred factual route appends memories rather than automatically revising or deleting existing facts. Explicit maintenance has separate APIs. Structural admission, exact-text filtering and nonempty-summary checks do not establish truthful extraction or faithful summarization. Expiry also differs by consumer: ordinary search can hide text that internal extraction still receives. These are static implementation findings, not observed downstream benefit. See Mem0's shared records, RTE-1, RTE-3, RTE-4, RTE-5, RTE-6 and BAP-1.
Dynamic Cheatsheet owns a different recurrence: cumulative mode solves, then replaces its sheet; retrieval-synthesis selects previous examples and constructs guidance before the next solution. The generator and curator receive selected context automatically. Whole-sheet replacement is wired, while requested semantic maintenance remains claimed. Benchmark targets score answers after memory assignment and cannot veto that assignment. A sampled retained artifact records answer 15 against target 315, stores the answer-derived rectangle strategy, and includes the previous sheet in the next generator prompt. This establishes observed artifact contents and recorded delivery. It does not establish that recalled content caused the next answer, or that the generating execution used the inspected revision. See Dynamic Cheatsheet's shared records, RTE-1, RTE-2, RTE-4, RTE-8, CLM-3 and ABS-1.
The comparison therefore supports named mechanisms and evidence limits. Automatic trace-fed writes with later consumers are wired in Mem0 and Dynamic Cheatsheet and afforded by Napkin's skill route; this classification does not establish improved capacity. No comparative causal benefit is established. No prevalence, performance ranking, trend, or Commonplace recommendation is inferred. Mem0's learning horizon and timing remain not-determinable; Napkin's mixed unions often have afforded basis; Dynamic Cheatsheet's curation union has claimed basis. Unknown assessments cannot become negatives. Derived feature counts and their exclusions are retained separately in the query ledger. A full refresh needs separately verified population coverage and consistent field interpretation.
Exact identities follow. Each run names the linked retained result.md; each review is kb/agentic-systems/reviews/<slug>.md.
| Run | Source revision | Result SHA-256 | Review SHA-256 |
|---|---|---|---|
| AAS-2026-09-26-napkin-01 | 7582d6a46f5a11995956e60a59c41a5b242109f1 |
c6be0f355181f46e0a53e1783145a0715b2bc136c8a0709661db74c2bb272c01 |
0362eec07dfe8adbc533b78dd5e7c67af33b301d31b6fe1d445f6819f84b83aa |
| AAS-2026-09-26-mem0-01 | 94c3fe9f238f3dbf29c9ce98643bd71eb13077cd |
1a540770af76846d221f95d02670989e2d41cb28003fd15358d345c5f9aad20a |
825719280a4b3960be992dd643450be4ee1a48d6a0763b94f787ed582ed49ee0 |
| AAS-2026-09-26-dynamic-cheatsheet-01 | 5cfe3c37e8e52b1d858d0f3df46e7f17c50991b9 |
29a349e828a7c17d301d85c872f644f86f530f6af9c0bdf4ff625277e2288c04 |
79b86f4947eabf4c83c5e64ee62144828ac84857d4035fd334a3afd3678e76c5 |
The temporary immutable input bundle is identified by manifest SHA-256 844f158f87906823765feb7371be95afc54929f3bdb6c28cb71aecb7c3d3bb82 and matrix SHA-256 d0901f96a8f43605f43e63a97a3f7a1eea21426c74244e96786ca3f17e2a7f95. It records checkout revision 26fbe60caa5431b958a7f01569b404ffa1978bc9; that revision alone is not asserted to retain every input byte. This is a workshop trial supported by a temporary bundle; a durable public synthesis would require retaining its complete evidence. The acceptance record records the local bundle location and the quantitative checks.
Verification mode: writer self-check, not independent review. Full bundled results, canonical records, source evidence, reconciliation and limits were read. Identity hashes were recomputed; current-source bundle verification passed with the expected manifest hash and --source-root . immediately before writing. No legacy review, transfer scan or newly acquired source supplied a finding.