llm-context-base synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
Prepare a file-capable assistant, not an internal runtime
llm-context-base is a personal knowledge-base template. It retains captures, project thinking, decisions, personal context and operating guidance in files, then instructs an external assistant to read, use and maintain them. This analysis covers the shipped substrate at Git commit 6d01cba8e2c22f9ca2519c70073a05cd54378a8c, with cutoff 2026-10-14. The whole-system boundary includes its prompts, templates, configuration, maintenance utility and integration recipes, not the external hosts that execute them. SRC-1 supplies predominantly doctrine/design evidence; the Python metadata converter supplies a narrow implementation layer. No observed target run or causal experiment is supplied (boundary and evidence, Source register).
The repository does not own inference, model selection, actual filesystem execution or permission enforcement. The foreground external assistant interprets natural-language protocols and owns sequencing; the host supplies tools and context handling. No exact model is selected, and provider parameter changes remain uninspected. Committed host permissions are configuration inputs, not a repository-enforced sandbox (RT-CMP-ai-assistant, RT-CLM-host-permissions).
In an ordinary initialized personal instance, a discovered bootstrap directs the assistant to configuration, general instructions and personal context. Phase-dependent loading reduces proactive questions and omits some introductory material over time; task routing loads additional modules when needed. Framework mode instead bypasses personal initialization and prohibits personal retention. Claude setup questions have an additional startup exception. These are prescribed branches, not an observed state machine (RT-RTE-session-start).
Capture, decide disposition and preserve only selected inputs
The representative workflow is capturing a process, choosing whether to file it, then asking about it in a later invocation. The shipped protocols explicitly connect those stages, but no trace demonstrates their completion. The same foreground assistant performs the work; the consequential handoffs are user task input, host file operations and user filing/structure decisions, not separate-agent delegation (workflow account).
Ordinary agent capture creates a dated inbox draft with summary, tags, update date and attribution, then asks whether to file it or leave it parked. A parked draft is a delivered capture, not a completed knowledge article. Filing selects a destination and template, updates relevant navigation/logs and applies type-specific completion criteria. The knowledge checklist includes accuracy, metadata, final placement, active status and related links. These criteria are instructions interpreted by the assistant, not an executable write barrier or an independent factual oracle (RT-RTE-capture-file, EPI-RTE-completion-check, EPI-RTE-completion-status, EPI-RTE-completion-links).
Filing approval selects disposition; it does not establish review of every sentence. Explicit destinations, append-only structural logs and immediate rule capture have different controls. Independent editor/clipper admission and the converter add further paths outside ordinary inbox review. Consequently, human-reviewable admission is a policy objective rather than universal quarantine (RT-CLM-review-policy).
Working articles may replace fuller inputs. Original preservation is off by default; when enabled, the assistant judges which originals warrant retention. Preserved originals are immutable by policy, not filesystem enforcement. Attribution identifies origin without establishing truth or faithful extraction. Decision records are more explicit about reasons: they provide context, alternatives, rationale and later outcomes, but outcome measurement and entry timing are not instrumented by the template (MEM-OBJ-sources, MEM-OBJ-knowledge, MEM-OBJ-decision).
File storage and conditional Git persistence provide continuity. Ignored journal/private bodies and uncommitted changes do not automatically obtain version-history recovery. Partial filing recovery is unspecified; available Git revert is not a transaction or automatic rollback (RT-RTE-capture-file, MEM-OBJ-knowledge).
Reload selected material for answers and decisions
A later question initiates a separate invocation. The assistant routes by question type to a directory, reads its index, scans summaries/tags and selects one to three relevant files. Files over 500 lines may be reduced to their first major section. Delivery is a cited synthesis with paths/update dates and applicable freshness warnings; a search miss calls for an explicit missing-document response and an offer to capture it. Saving an answer is a further write decision, not an automatic query effect (RT-RTE-knowledge-query, workflow query annotation).
This mechanism limits loaded context through file selection rather than an inspected vector engine. Summaries are relevance proxies, not complete content. Unloaded files remain reachable on disk but are not live pointers in the model window: their text must be read again. Startup/task routing supplies learned context and current guidance without a separate retrieval request; ordinary question answering is requested retrieval. Decision conversations also trigger proactive selection of relevant completed outcomes and failure constraints. Those reports are evidence offered to deliberation, not expected-answer references or proof of a transferable causal mechanism (MEM-OBJ-metadata, MEM-OBJ-decision, memory query annotation).
The strongest supported decision-learning contribution is this prescribed reuse of rationale and outcomes. Whether it improves future decision capacity remains uninspected at the template boundary. The public experience-learning claim exceeds demonstrated benefit, not the existence of its prescribed consumer path (EPI-CLM-decision-learning).
Optional manual-compaction hooks supply a different context-control mechanism: a sentinel triggers reinitialization on the next prompt. This affords rereading durable files, not restoration of discarded turns or recovery of unsaved discoveries. Host summary contents and hook deployment remain opaque. Linked-code documentation separately affords wiki-context consultation by an assistant in another repository; actual deployment and full-window delivery are unobserved (MEM-RTE-compact-reload, MEM-BAP-linked-context). Other external-consumer recipes are incompletely inventoried, as the report verification explicitly records (report-verification limits).
Retain conversational discoveries and revise operating guidance
The main adaptation contribution is a prescribed conversation-to-durable-text chain. Clear rules, preferences and usage observations are selected into configuration, personal context, instructions, templates or structure. Future sessions and tasks are instructed to consume the changed files. Static shipped guidance becomes retained behavioral memory when use revises it. This is textual adaptation, not established weight training; reasons and original turns need not accompany every retained rule (MEM-RTE-pattern-capture, MEM-OBJ-config, MEM-OBJ-context, RT-BAP-instruction-load).
Admission depends on the branch. Clear established rules request immediate writes; ambiguous capture asks the user. New directories require approval. Instruction optimization separately requires approval before restructuring. It diagnoses large/repeated reads, proposes splitting, conditional loading or compaction, and retains approved text changes, a review date and log. Line counts and perceived slowness motivate review but do not demonstrate faster or more capable later action (MEM-RTE-pattern-capture, RT-RTE-context-review).
The accepted trace-learning classification recognizes these automatic conversation-fed durable writes at the claimed layer. It does not assert observed activation or improved capacity. The profile verification preserves unresolved outcome, saved-synthesis and operation-summary chains rather than declaring them absent (profile verification).
The theory and improvement properties remain independent, at the personal-instance operating-text boundary and intended horizon of later sessions:
- Theory-builder condition 1 — localized content: afforded. Whole written rules/guidance units are inspectable and revisable.
- Condition 2 — consumption: claimed. Later decisions are prescribed to apply retained guidance; actual dependence is unobserved.
- Condition 3 — content-directed criticism: uninspected. No supplied episode establishes a challenged explanatory claim, blame placement and resulting revision or changed reliance.
- Condition 4 — iteration: claimed. Retained edits and later correction/review support a prescribed next round, not observed criticism-fed iteration.
- Criticism improving future-action capacity: uninspected. Neither revision nor persistence establishes this learning result.
- Reflection: a representation-mediated feedback loop is prescribed for selected usage and instruction organization; exercised two-way causal dependence remains uninspected. Reflective theory-builder status is likewise uninspected.
- Autonomy: autonomous theory-builder status is uninspected; explicit human approval contributes to some internal selections. This does not negate automatic writes on other branches.
- Self-improvement: a claimed dispositional, improvement-directed path targets adaptation to user needs and lean useful context. Evidence-responsive determination, operative installation and subsequent causal dependence are not demonstrated by a run. Exercised and successful self-improvement remain uninspected.
These statuses come from the separately bounded pattern-capture and optimization accounts, not a system-wide grade (MEM-RTE-pattern-capture, RT-RTE-context-review).
Maintain structure without inferring factual acceptance
Lint requests structural/age checks and cleanup; archive/supersede retains reasons and usually reduces retrieval priority rather than universally blocking consumption. Completion names accuracy but supplies no demonstrated evidence-consuming acceptance of a newly produced ampliative claim. Query synthesis has no concrete input/output witness establishing reshaping, entailment or non-entailed conjecture. Blog review addresses presentation readiness; publication gives content operational exposure without adding factual warrant (MEM-RTE-lint-maintenance, EPI-RTE-lint-check, EPI-RTE-blog-review, EPI-RTE-blog-publish, epistemic conclusion).
The executable exception is bulk metadata conversion. Fixed code selects matching files and overwrites them without per-file approval, backup or transaction. Static analysis establishes possible pre-title loss and duplicate-key replacement; a converted label proves neither YAML validity nor semantic preservation (RT-RTE-metadata-migrate, EPI-RTE-metadata-transform).
For the documented scenario of modest file-based capture, selective answering and personalized continuity, the template provides a coherent prescribed progression and editable durable state. For a scenario requiring universally reviewed admission, complete retrieval, factual acceptance, privacy enforcement or demonstrated improvement, these records do not establish the required guarantees. Aligned templates/metadata/link conventions would remove specific contract ambiguities. Authentic write-to-later-consumer traces would establish activation; evidence-consuming acceptance records would support bounded epistemic conclusions; comparable before/after tasks with controlled interventions would be needed to attribute improvement. Installed integration records would change deployment findings without by themselves proving benefit.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| Contradictory original-disposal timing; carried from report verification | SRC-1; MEM-OBJ-inbox, MEM-OBJ-sources, RT-RTE-capture-file | Source/inbox READMEs and write protocol at the pin | Uniform disposal timing or automatic seven-day TTL; optional preservation and filing-conditioned disposal remain prescribed | Reconciled timing instructions and, for operation, a trace of the selected disposition |
| Metadata/link conventions disagree; carried from report verification | SRC-1; MEM-OBJ-metadata, MEM-RTE-lint-maintenance, RT-RTE-knowledge-query, EPI-RTE-lint-check | Standard, query and lint instructions | Consistent metadata matching, complete schema enforcement or coherent orphan detection; individually specified predicates remain supported | Aligned formats and checks covering required fields/link forms, with inspected validation results |
| Active template versus draft admission; specialized published status outside canonical lifecycle; carried from report verification | SRC-1; EPI-RTE-completion-status, EPI-OBJ-knowledge-article, MEM-OBJ-knowledge, workflow account | Knowledge template, write/completion protocols and blog lifecycle | Uniform template-to-status contract or inference that copied/published status establishes completion or factual acceptance | Consistent lifecycle/template rules and a supported completion/acceptance episode |
| Additional external-consumer recipes lack individual route records; carried from report verification | SRC-1; RT-CMP-ai-assistant, RT-RTE-knowledge-query, MEM-BAP-linked-context, memory read-back, verification finding | Supported-tools recipes for Claude Projects background files, on-demand ChatGPT access and push-triggered external vector-store sync | Complete external-consumer inventory or cross-host selector, persistence, embedding, warrant/privacy comparisons; traced foreground reads remain valid | Separate producer/selector/persistence/consumer records and inspected external implementation/deployment evidence |
| No authentic operation, corpus or intervention evidence | SRC-1; RT-RTE-capture-file, RT-RTE-knowledge-query, MEM-OBJ-decision, MEM-RTE-pattern-capture, RT-RTE-context-review | Frozen template, placeholders/examples and prescribed loops | Observed delivery, retrieval quality, decision dependence, exercised reflection/theory criticism/self-improvement or causal benefit | Authentic episode and later-consumer traces; objective-specific outcomes and controlled comparisons for attribution |
| External execution, permissions and compaction are not inspected implementations | SRC-1; RT-CMP-ai-assistant, RT-CLM-host-permissions, MEM-RTE-compact-reload, MEM-OBJ-knowledge | Host/configuration, compaction and journal-sync recipes only | Reliable adherence, host security/privacy, model mutability, hidden summary lineage or restored unsaved discoveries | Pinned host/integration implementations, effective permissions, installed hooks and inspectable input/output payloads |
| Optional sources and selective reads leave lineage and received content incomplete | SRC-1; MEM-OBJ-sources, MEM-OBJ-metadata, RT-RTE-knowledge-query | Default source preservation off; summaries and 1–3-file/long-file selection | Guaranteed extraction reconstruction, complete evidence delivery or warrant preservation | Retained originals with transformation provenance and actual selected/delivered content traces |
| Acceptance and additional semantic/admission chains remain unresolved | SRC-1; EPI-RTE-completion-check, MEM-OBJ-decision, RT-RTE-knowledge-query, EPI-RTE-blog-publish, MEM-OBJ-logs, MEM-OBJ-sources | Accuracy checklist; outcome/save/publication/log/original-reprocessing roles; no concrete semantic pairs | Accepted newly produced ampliative knowledge; complete write-agency, curation, trace-origin or external read-back classification | Candidate/input/output, evaluator evidence, criterion/use/scope, admission decision and subsequent consumer records |
| Mechanical conversion is not validated or transactional | SRC-1; RT-RTE-metadata-migrate, EPI-RTE-metadata-transform | Complete static converter account and loss witness; no execution | Lossless conversion, valid YAML for all consumers or automatic recovery from partial failure | Restricted checked input/consumer domain, preservation/validation guarantees and recoverable execution evidence |
No unresolved cross-report conflict was retained. Report verification declares the first four limits; profile verification declares none. No frozen-source files were read for this synthesis; it uses the supplied accepted members without adding records or evidence.