Napkin synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
Napkin supplies persistent files and callable retrieval to agents; it is not the agent runtime that decides what to do next. This account covers its shipped CLI/SDK, scaffolding, packaged distill/tend skills, design documents and benchmark callers at commit 7582d6a46f5a11995956e60a59c41a5b242109f1, with an analysis cutoff of 2026-10-12. The evidence is source implementation, instructions and attributed benchmark reports from SRC-1, not observed execution or causal experiments. External pi integration, providers, dependency internals, datasets and run results are not frozen here. The Source register and boundary distinguish these access gaps from available code left partly uninspected.
Establish a vault, then disclose retained material
A CLI wrapper constructs the SDK, resolves the vault, invokes core functions and returns text or JSON. Discovery walks upward for supported vault layouts; if none exists, even a retrieval invocation can initialize a bare vault. The human or external agent owns the next request. Core operations execute as the local process and use its filesystem privileges; choosing a vault path is not a sandbox or permission grant (RT-CMP-napkin, RT-OBJ-vault-info, RT-RTE-vault-retrieval).
The ordinary progression is overview → search → full read. Overview supplies a folder map, descriptions and search handles, plus the whole nonempty trimmed NAPKIN.md. Search ranks lexical matches with backlink and relative-recency contributions, truncates the hit list and returns snippets. Full read returns an untruncated file. These are explicit requests for retained material, not verified unsolicited model-context injection. Direct reads bypass the staged progression. Per-operation limits reduce exposure, but there is no shared model-token budget; full context and individual matching lines can remain large (RT-RTE-vault-retrieval, MEM-OBJ-napkin, memory read-back).
The retained state includes natural-language note bodies and context, symbolic frontmatter/tasks/link targets and access metadata, persisted templates, configuration and derived caches. Base definitions, canvas JSON and bookmarks have separate requested consumers; a canvas or bookmark reference does not itself fetch its destination. The base query database is transient, not a persistent SQL memory store (MEM-OBJ-markdown-notes, MEM-OBJ-note-templates, MEM-OBJ-base, MEM-RTE-structured-access).
Search and overview reuse file-backed projections under path/mtime fingerprint checks. Search can rank through an older index while rereading current bodies for snippets. Preserved mtimes can conceal content changes, and folder-scoped fingerprints omit backlink changes originating outside that folder. Overview probes candidate handles against search, but failed-probe fallback and roster completion prevent a universal top-three retrieval guarantee. These mechanisms concern access and freshness, not endorsement of the notes (MEM-OBJ-search-cache, MEM-OBJ-overview-cache, RT-CLM-keyword-probes).
Turn a session into future-use notes
Napkin's strongest supported contribution beyond retrieval is a packaged procedure for selecting session findings and integrating them into durable, topic-based notes. Distill asks an external agent to retain investigated fixes, confirmed behavior, decisions and their reasons; skip routine material; apply a three-month usefulness test; and mark drawn generalizations. It treats conversation and quoted sources as data rather than instructions. This is claimed trace-fed retention and consolidation with an intended later reader, not demonstrated improved capacity for future action (RT-RTE-distill-session, EPI-CLM-distill-knowledge, profile verification).
The representative workflow connects those choices to delivery. The same host agent groups topics, requests overview and a five-hit duplicate search, reads a same-subject note, then merges or creates. No hit in that bounded search does not prove that no overlap exists. A merge should integrate information into existing sections and explicitly handle contradictions. The model decides the content; CLI/SDK calls persist it. User initiation is not per-note approval, and the core's existing-target create guard is not semantic acceptance (workflow account, RT-RTE-caller-mutation).
The new-note example has a consequential instruction/executor mismatch. Its folder-valued --path is interpreted as a complete filename, while template selection copies template bytes rather than populating distilled knowledge. Literal execution therefore cannot certify the intended titled, populated note. After writing, the agent is asked to check unresolved links and repair them, then report one line per written note. That diagnostic does not stage or roll back the preceding writes. It tests recognized wikilink target resolution, not claims, intended duplicate referents or heading validity; default output lacks referring-file provenance (workflow mutation annotation, MEM-OBJ-note-templates, WF-RTE-distill-verification).
Tend supplies a separate conservative upkeep policy: select at most 3–5 issues, defer uncertain merges, repair links, integrate obvious duplicates and retire notes to trash. A proposed new schema remains a user decision. These bounds and the data/instruction rule depend on host/model interpretation, not core enforcement. Raw writers and unrestricted APIs bypass them. Ordinary trash deletion withdraws notes from walker-based search/listing but leaves exact-path reads possible; permanent deletion unlinks the file. Overwrites lack a separate archive, and caller repair is not transactional recovery (RT-RTE-tend-vault, RT-BAP-skill-guidance, MEM-OBJ-markdown-notes, RT-RTE-caller-mutation).
The separately documented background design uses a cheap model alongside the working agent to capture new conversation entries. Its scheduler, parser, cursor persistence and admission enforcement are outside the frozen implementation. It must not be combined with the packaged skill into a wired automatic service. Similarly, calling NAPKIN.md always-loaded supports a claimed broad host-delivery path, not verified injection (MEM-RTE-background-distill, MEM-OBJ-napkin).
Answer questions and interpret checks within their domains
The shipped benchmarks exercise a different intended route: deterministic dataset ingestion followed by requests to an external pi answer agent, with skills disabled. LongMemEval writes raw conversational rounds and date-derived mtimes into a per-question vault. LoCoMo imports dialogue plus supplied summaries/observations and can reuse one sample vault across questions. HotpotQA writes supplied paragraphs and heuristic links. These source-supported later-consumer paths do not demonstrate actual reads or successful answers, and they do not test distill/tend compliance (RT-RTE-benchmark-answer, MEM-RTE-benchmark-variants, MEM-OBJ-session-summary).
Checks have different licenses. Link diagnostics establish parser/resolver results. Numeric base summaries establish conditional arithmetic over encoded inputs, not source-premise truth or arbitrary formula fidelity. Neither stored status labels nor note retention establish evidence-consuming acceptance of a novel claim. Distill's inferred marker leaves the semantic relation unresolved without concrete inputs, outputs and judgments (EPI-RTE-link-check, EPI-RTE-base-summary, EPI-RTE-property-edit, epistemic conclusion).
Benchmark scores affect reporting, not an inspected successor-note or policy update. LongMemEval can grant full credit by substring before model judging; judge failure falls back to token F1. LoCoMo/HotpotQA preprocessing can remove answer qualifications under gold-conditioned length rules, and token overlap ignores word order. Access metrics are string proxies, not proof of complete reads. Reported headline accuracy remains attributed operation, not independently verified correctness or causal retrieval benefit (EPI-RTE-longmem-judge, EPI-RTE-answer-extract, EPI-RTE-token-score, EPI-RTE-score-report, EPI-CLM-benchmark-results).
Assessment at the supported horizon
For caller-directed reuse of retained files, persistence and requested retrieval are wired. For turning working-session experience into future guidance, the contribution is a claimed semantic procedure with inspectable command effects; successful delivery, faithful transformation and improved future capacity remain unresolved. For scheduled unattended upkeep, conservative instructions exist, but reliable compliance and safe recovery are unverified. For temporary benchmark answering, ingestion, external-agent requests and scoring are wired, while exercised dependence and component benefit are unobserved. These scenario differences rule out one system-wide epistemic grade (RT-RTE-vault-retrieval, RT-RTE-distill-session, RT-RTE-tend-vault, RT-RTE-benchmark-answer).
The accepted runtime records do not provide a formulated-theory assessment establishing theory-builder conditions 1–4. Retained reasons and contradiction instructions must not stand in for consumed theory, content-directed criticism or iteration. Criticism of a consumed theory improving future capacity is not established. Reflection and autonomous theory-building are independently unestablished; external scheduling doctrine does not establish either. Exercised self-improvement is also unestablished: no supplied subsequent operation causally depends on an evidence-responsive change to the bounded system's own organization. These are withheld attributions, not evidence that the properties are absent in uninspected host behavior (RT-RTE-distill-session, RT-RTE-machinery-change, epistemic conclusion).
Concrete changes would alter these assessments: corrected create instructions and a populated-note delivery trace would address the workflow gap; pinned host/extension code and traces would resolve automatic capture and injection; candidate-specific evidence and disposition records would support acceptance findings; and comparative runs with independent component variation would support retrieval-benefit attribution. A retained-update trace followed by causally dependent later behavior would be needed for exercised self-improvement, with separate outcomes for successful improvement. No source was read beyond the handed members for this synthesis.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| Distill create example misstates placement and leaves template content unpopulated; carried from report verification | SRC-1; RT-RTE-caller-mutation, MEM-OBJ-note-templates, report limits, workflow annotation | Shipped skill example and CLI/core create semantics at the reviewed commit | Successful end-to-end new-note delivery from literal execution of that example | Correct file-path/population sequence and an inspectable trace showing titled, populated retained notes |
| Overview exposure scorer has a string-argument/VaultInfo signature mismatch; carried from report verification | SRC-1; RT-CLM-keyword-probes, memory integration issues, report limits | Scorer caller and search signature at the reviewed commit; overview is a separate supported path | Scorer-based keyword-quality or exposure conclusions | Corrected caller plus inspectable evaluation inputs and results; universal handle validation would also need the fallback exceptions addressed |
| Auxiliary semantic inventories remain partial; carried from report verification | SRC-1; MEM-RTE-structured-access, EPI-RTE-base-summary, memory coverage, epistemic coverage, report limits | Partly inspected base formula/filter, overview tokenizer and graph-renderer internals; available source, not access gaps | Exhaustive derivation/evaluator, UI-consumption and selector-quality comparisons; system-complete epistemic negatives | Completed source tracing of the relevant internals and their consequential consumers |
| No observed runs, datasets or raw benchmark results supplied | SRC-1; RT-RTE-benchmark-answer, MEM-RTE-benchmark-variants, EPI-CLM-benchmark-results, Source register | Source drivers, prompts and reported tables only; no target execution | Reproduced accuracy, actual retrieval reliance, skill compliance, learning benefit or causal component efficacy | Pinned datasets and complete execution/results evidence; controlled independent variation for component attribution |
| External host, extension, provider and dependency behavior is unavailable | SRC-1; MEM-RTE-background-distill, MEM-OBJ-napkin, MEM-OBJ-search-cache, RT-CMP-pi-model | Napkin's calls and doctrine; external implementation not frozen | Wired automatic scheduling/injection, cursor format and persistence, full index-payload classification, exact model weights or parameter-update conclusions | Frozen integration/dependency/provider evidence covering the unresolved part, with run traces for operation claims |
| Initial provenance and candidate-level semantic warrant remain unresolved | SRC-1; MEM-OBJ-markdown-notes, MEM-OBJ-session-summary, EPI-OBJ-vault-prose, RT-RTE-distill-session | Generic supplied files, dataset-summary copying and instructed session transformation | Complete lineage/admission classification, faithful concrete derivation, accepted ampliative production, or criticism-led improved future capacity | Content provenance and concrete session-to-note pairs with evidence-consuming decisions, retained criticism and later-consumer evidence appropriate to each claim |