MemWiki epistemic report

Type: agentic-system-analyses/types/agentic-system-epistemic-report.md

Source-and-claim boundary

The question is how MemWiki turns project material into retained assertions, checks those assertions and licenses later reliance. Scope and exclusions: see boundary. Evidence identity and layers: see the Source register, SRC-1. This report directly inspected the source paths below at the registered revision; runtime findings are supplied context, not substituted for those reads. No target execution occurred. No new source or evidence layer is introduced.

MemWiki's executable installs text; its continuing epistemic routes are instructions to an external agent. The relevant public claims are EPI-CLM-project-memory, EPI-CLM-primary-truth, EPI-CLM-health-fixes and RT-CLM-safe-update. The model component remains opaque: see RT-CMP-host-agent. Missing host implementation prevents conclusions about actual discovery, compliance, approvals or weight changes. Missing raw inputs and model traces prevent reconstruction of individual synthesis edges. Missing execution comparisons prevent benefit and causal attribution. Missing npm artifact prevents installed-byte equivalence.

Entry point or operation	Source path directly read, SRC-1	Covering records or bounded exclusion
Declared CLI and build selection	package.json, src/index.ts	RT-RTE-cli-init, RT-RTE-protocol-update; build configuration supplied through runtime, not directly reread
Alternate executable	src/index.js	RT-RTE-legacy-init; init only, older protocol
Host hooks	.cursorrules, CLAUDE.md, .github/copilot-instructions.md	RT-BAP-host-protocol, EPI-CLM-primary-truth
Raw-source reading, discoveries, hot/log writes, page/index organization	AGENTS.md, src/index.ts	RT-RTE-agent-session
Repository ingestion	AGENTS.md, README.md	RT-RTE-agent-commands
Health checking and proposed fixes	AGENTS.md, README.md	EPI-RTE-wiki-lint
History condensation	AGENTS.md, README.md	EPI-RTE-log-fold
Retained statements and routing	.memory/wiki/hot.md, .memory/wiki/index.md, .memory/wiki/stack.md, .memory/wiki/patterns.md, .memory/wiki/bugs.md, .memory/wiki/decisions.md, .memory/wiki/log.md	EPI-OBJ-hot, EPI-OBJ-domain-pages, EPI-OBJ-log, EPI-OBJ-index
External model/tool implementations and published distribution	Boundary exclusions	No claim about their evaluators or runtime behavior; prevents deployment-complete conclusions

CLI dispatch admits only init and update; there is no registered executable slash-command tool in either inspected executable. Lint and fold are nevertheless material shipped operations because the installed protocol requests them. Available protocol and maintenance paths were assessed, including the strongest counterexample to a blanket “no checking” claim: lint explicitly requests a health evaluator. The absence conclusion below concerns an explicit acceptance decision, not all possible model checking.

Epistemic-object inventory

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source ID and local anchor	Gap/limit
RT-OBJ-protocol-files	Mostly obligations, not truth-apt updates; project-memory assertions also occur	see record	SRC-1, AGENTS.md, CLAUDE.md	Instructions confer requested priority, not warranted truth
EPI-OBJ-hot	Project initialized; ingestion completed; next steps are policy/task content	see record	SRC-1, .memory/wiki/hot.md	Ingestion success is reported, not independently checked
EPI-OBJ-domain-pages	Stack facts, CLI patterns, decisions and rationale; bugs page has no bug instance	see record	SRC-1, .memory/wiki/stack.md, .memory/wiki/patterns.md, .memory/wiki/decisions.md, .memory/wiki/bugs.md	Assertions are available; production and acceptance provenance unknown
EPI-OBJ-log	Work and ingestion assertions	see record	SRC-1, .memory/wiki/log.md	Historical report is not an execution trace
EPI-OBJ-index	Page descriptions and links	see record	SRC-1, .memory/wiki/index.md	Routing is not evidence validation
Raw documents	Potential external facts; concrete content not determinable	Source of truth in README	SRC-1, README.md, AGENTS.md	No tracked raw example; cannot assess source truth or a concrete import
RT-CMP-host-agent	Parametric knowledge not individuated	see record	SRC-1, README.md	Model version, endpoint and parameter changes remain uninspected

Authority-route ledger

Fields owned by cited records say see; the local classifications below assess separate consequential functions, not replacement route identities. For natural-language routes, “doctrine only” means the requested action is present, not that the host reliably performs it.

Route ID	Route function	Architectural status	Object/candidate ID	Content/update relation	Transition or check target	Evaluator/condition and domain	Activation and timing	Possible or observed result	Implemented force	Epistemic authority and scope	Operational authority: behavior permitted, blocked, or changed	Behavioral-authority path: consumer, channel, force, horizon	Evidence source ID and local anchor	Claim IDs or none	Mismatch marker or none	Gap/limit
RT-RTE-cli-init, RT-RTE-legacy-init	content transformation	implemented	Protocol and starter pages	truth-apt transformation: acquisition/import	Fixed template to file	Destination existence, filesystem domain	see records	see records	Exact UTF-8 write or skip, not truth check	Source text preserved; source warrant unknown	Installs requested future instructions and starter assertions	see records and RT-BAP-host-protocol	SRC-1, src/index.ts, src/index.js	EPI-CLM-project-memory	none	Starter assertion can be false if files predate current invocation
RT-RTE-protocol-update	behavior/policy adaptation	implemented	RT-OBJ-protocol-files	non-truth-apt policy/content update: fixed instruction replacement	Local hook/protocol replacement	Exact path predicate, not evidence evaluation	see record	see record	Unconditional four-path overwrite	No license to treat local wiki as correct	Can replace custom host guidance	see record and RT-BAP-host-protocol	SRC-1, src/index.ts	RT-CLM-safe-update	Missing wiki paths still created	Downstream behavior unobserved
RT-RTE-agent-session	content transformation	doctrine only	Raw documents	truth-apt transformation: acquisition/import	New external material enters host context	Agent reading; source-specific domain unknown	see record	see record	Requested reading; no source authentication	Raw file preservation requested; source warrant unknown	Supplies potential premises for work/synthesis	see record and RT-BAP-host-protocol	SRC-1, AGENTS.md, README.md	EPI-CLM-primary-truth	Preservation is not endorsement	No raw instance or read trace
RT-RTE-agent-session, RT-RTE-agent-commands	content transformation	doctrine only	EPI-OBJ-domain-pages, EPI-OBJ-hot, EPI-OBJ-log	truth-apt transformation: indeterminate	Raw/code/task findings become prose	Host model; project facts and selected rationale	see records	see records	Requested synthesis/write	No entailed derivation or accepted ampliative claim established	Model may author later context without per-sentence approval	see records and RT-BAP-host-protocol	SRC-1, AGENTS.md	EPI-CLM-project-memory	none	No paired inputs/outputs; reshaping, derivation and conjecture remain unresolved
RT-RTE-agent-session	retention	doctrine only	Hot, domain pages and log	no content change	Keep newly written content for later invocation	Host file tools; no acceptance criterion	see record	see record	Requested persistence	Retained candidate, not accepted knowledge	Makes authored content available later	see record and RT-BAP-host-protocol	SRC-1, AGENTS.md	EPI-CLM-project-memory	none	Existing pages show candidates available, not route activation
RT-RTE-agent-session	operational admission/selection/consumption	doctrine only	Hot, index and selected domain content	no content change	Use available wiki as project context	Host model judges whether hot is sufficient; not a truth evaluator	see record	see record	Requested primary-source priority and conditional reads	Operational reliance requested without explicit acceptance	Context may affect answers, changes and next steps	see RT-BAP-host-protocol	SRC-1, AGENTS.md, CLAUDE.md	EPI-CLM-primary-truth	Priority exceeds demonstrated warrant	No trace of which content reached or changed a decision
EPI-RTE-wiki-lint	check/evidence production	doctrine only	Wiki, especially hot and empty domain pages	no content change	Staleness, missing context, empty pages	see record	see record	see record	Requested diagnosis/proposal, not enforced invalidation	At most model diagnosis in project-health domain	Proposal can guide later edits; no automatic exclusion	see record	SRC-1, AGENTS.md	EPI-CLM-health-fixes	README says fixing, protocol says propose	No diagnosis instance or accuracy evidence
EPI-RTE-log-fold	content transformation	doctrine only	EPI-OBJ-log	truth-apt transformation: indeterminate	Older history to summary	see record	see record	see record	Requested replacement with crucial-history preservation	Intended compression; fidelity not checked	Shorter history can become later context	see record	SRC-1, AGENTS.md	EPI-CLM-project-memory	Append-only description versus condensation	No paired log/summary; cannot assign semantic preservation

No disposition/acceptance or post-acceptance lifecycle-integration route was found within the inspected protocol/CLI boundary: see EPI-ABS-acceptance-gate. Linking a new page changes organization for later access, but does not establish lifecycle integration as defined here because acceptance has not been established. Lint's possible detection is not merged with an invented rejection mechanism.

Concrete deterministic trace, deduced from implementation without execution: take argv command update, an existing CLAUDE.md containing arbitrary custom text, and an existing .memory/wiki/hot.md containing a stale statement. The exact-path predicate makes the first destination a protocol file; its old content is replaced by the fixed hook template. Hot is not a protocol file and exists, so its stale text is skipped unchanged. Thus mechanical update neither validates nor refreshes that candidate. For command lint, dispatch exits 1 before any template loop. SRC-1, src/index.ts; supporting branches are retained in RT-RTE-protocol-update and RT-RTE-cli-init. This is an implementation deduction, not observed operation.

System-claim versus route comparison

Claim ID	Claimed operation or warrant	Claim source ID/anchor and evidence layer	Doctrine/design support	Implemented route IDs	Observed-run support	Causal support and design limits	Supported conclusion	Mismatch/unknown
EPI-CLM-project-memory	see record	SRC-1, README.md, AGENTS.md; doctrine/design	see record	RT-RTE-cli-init; continuing loop RT-RTE-agent-session is doctrine only	None; log reports ingestion	No contrast; external compliance required	Files and requested cross-session reuse, not demonstrated compounding benefit	Retention does not establish learning or truth
EPI-CLM-primary-truth	see record	SRC-1, CLAUDE.md, README.md; doctrine/design	see record	Hook installation; RT-BAP-host-protocol requests reliance	None	No evidence-consuming acceptance gate	Project-priority instruction exists	Domain facts can contradict metadata; source warrant unknown
EPI-CLM-health-fixes	see record	SRC-1, README.md, AGENTS.md; doctrine/design	see record	EPI-RTE-wiki-lint is doctrine only	None	No freshness oracle, validity criterion or enforced withdrawal	Requested health assessment and proposals	Automatic fixes stronger than protocol text
RT-CLM-safe-update	see record	SRC-1, README.md, src/index.ts; doctrine/design and implementation	see record	RT-RTE-protocol-update	None	Ordinary filesystem assumptions; no concurrent-writer protection	Existing wiki files skipped	Missing starter pages created; old facts preserved regardless of truth

Bounded conclusion

MemWiki implements file installation and fixed protocol replacement. It requests acquisition of source material, synthesis, retention, selective reuse, health checking and history condensation. These are distinct from observed performance. Source files are intended to remain unchanged, but there is no demonstrated lineage from individual source assertions to wiki sentences. Import warrant remains unknown; untouched source bytes alone do not establish truth. Source-dependent claims: SRC-1, src/index.ts, AGENTS.md, README.md; see the ledger records.

The strongest evaluator is the requested lint pass over staleness, missing hot context and empty domain pages. Its result is a proposal, with advisory force through the external agent/user; no required consumer marks a candidate accepted or withdraws it from later reading. Retained wiki is requested as project truth before such checking. No warranted entailed derivation, evidence-consuming acceptance of an ampliative claim or subsequent lifecycle integration is established in this inspected boundary. This does not deny that a host can test code or critically inspect facts during ordinary work. Those external actions are not supplied evidence of MemWiki acceptance. SRC-1, AGENTS.md, CLAUDE.md; EPI-RTE-wiki-lint, EPI-ABS-acceptance-gate.

A concrete contradiction qualifies the warrant claim: stack.md asserts no runtime dependencies, while package.json declares memwiki. This demonstrates a stale or mis-scoped candidate assertion at the frozen commit, not a known failure of lint execution. Similarly, the log reports links to landscape.md and competitors.md, while the retained index contains neither; the report cannot reconstruct when or why those references diverged. Neither inconsistency licenses a system-wide verdict. SRC-1, .memory/wiki/stack.md, package.json, .memory/wiki/log.md, .memory/wiki/index.md; EPI-OBJ-domain-pages, EPI-OBJ-log.

Instruction replacement can change future policy delivery without a truth-apt update. Retained work summaries and rationale could guide later action, but exercised learning, reflection, autonomous maintenance, self-improvement and a theory-criticism iteration remain independently unestablished. External model parameters and internal reasoning are uninspected; that opacity establishes neither those properties nor their absence. SRC-1, src/index.ts, AGENTS.md; RT-CMP-host-agent, RT-RTE-agent-session.

Shared records

Operative objects

EPI-OBJ-hot

Label: Hot cache

Source-native identity: .memory/wiki/hot.md. Natural-language project state plus task checklist, stored in Markdown on disk. Distinct from RT-OBJ-protocol-files: this is retained context the hooks request, not the hooks themselves. SRC-1, .memory/wiki/hot.md, AGENTS.md. No supplied operative-object counterpart describes this page.

• Successfully completed /memwiki-ingest, populating the wiki with stack details, coding patterns, and architectural decisions. --- .memory/wiki/hot.md

EPI-OBJ-domain-pages

Label: Domain pages

Source-native identity: .memory/wiki/stack.md, .memory/wiki/patterns.md, .memory/wiki/decisions.md, .memory/wiki/bugs.md. Natural-language facts, conventions, rationale and bug-record placeholder in Markdown files. These share the session write mechanism and task-dependent consumer, unlike the first-read hot cache and fold-targeted log. Distinct from RT-OBJ-protocol-files; no supplied counterpart describes these candidate pages. Source lineage and acceptance provenance are unknown. Some patterns accurately describe the inspected code; correctness of those instances does not establish correctness of the whole set. SRC-1, .memory/wiki/patterns.md, .memory/wiki/stack.md, .memory/wiki/decisions.md, .memory/wiki/bugs.md, package.json.

• Zero Dependencies: The project explicitly avoids runtime dependencies (only @types/node and typescript as devDependencies) to ensure zero friction and high security/stability. --- .memory/wiki/stack.md

"dependencies": { "memwiki": "^1.0.0" } --- package.json

• Decision: Store the .cursorrules, CLAUDE.md, and default wiki files as string literals directly in src/index.ts. --- .memory/wiki/decisions.md

EPI-OBJ-log

Label: Work log

Source-native identity: .memory/wiki/log.md. Natural-language historical assertions in a Markdown file. Unlike RT-OBJ-protocol-files, this contains reported work and is the specific fold input; no supplied object counterpart. It is intended for later history access, but actual causal use is unobserved. The claimed ingestion supplies reported operation only, not authenticatable route execution. SRC-1, .memory/wiki/log.md, AGENTS.md.

• Triggered /memwiki-ingest. Scanned package.json, tsconfig.json, src/index.ts, and README.md.

• Populated stack.md, patterns.md, and decisions.md based on project findings. --- .memory/wiki/log.md

• Updated index.md to link to the new landscape.md and competitors.md files. --- .memory/wiki/log.md

EPI-OBJ-index

Label: Project wiki index

Source-native identity: .memory/wiki/index.md. Natural-language labels and wiki-link notation in a Markdown file; links are interpreted by the external agent, with no shipped parser or checker. Distinct from RT-OBJ-protocol-files: it supplies conditional context selection rather than detailed maintenance rules. No supplied object counterpart. Its registered links expose the default pages, not the two added pages asserted by the retained log. SRC-1, .memory/wiki/index.md, .memory/wiki/log.md, AGENTS.md.

• [[hot.md]]: Immediate context and next steps.
• [[log.md]]: Append-only changelog of agent sessions.
• [[stack.md]]: Tech stack details and versions.
• [[patterns.md]]: Established coding patterns and conventions.
• [[bugs.md]]: Known issues, quirks, and workarounds.
• [[decisions.md]]: Architecture Decision Records (ADRs). --- .memory/wiki/index.md

Routes

EPI-RTE-wiki-lint

Label: Wiki health check Part of: RT-RTE-agent-commands

This is the health-check part, separated from ingestion and fold by target, result and force. Endpoints/progression: user chat command → host interpretation → wiki diagnosis → proposed fixes. Owner: external agent; state/context: selected project/wiki content; action effects: requested diagnosis and proposal, not enforced rejection. Check target/domain: outdated wiki assertions, missing immediate context and empty pages. Evaluator: host model; no supplied expected-answer reference, independent oracle or truth metric. User proposes the launch; model decides findings and proposals; human veto is host-dependent and uninspected. Mode: open maintenance request. Trigger: /memwiki-lint; guidance names those three health conditions. Proposed fixes persist only if separately acted on; no rollback, required approval or exclusion marker is supplied. SRC-1, AGENTS.md, README.md.

• implementation conclusion status: claimed
• operation conclusion status: uninspected
• Immediate return: Requested health findings and proposed fixes.
• Later read-back: uninspected — protocol does not require persisting diagnosis or applying fixes for later sessions.
• Delegated visibility: uninspected — host delegation is unspecified.
• Selection predicate: Chat token /memwiki-lint; host selects wiki evidence for the stated health conditions.
• Invalidation or expiry: No enforced expiry or withdrawal; outdated information is a requested diagnosis target.
• Activation or effect: Advisory proposal through host response; any subsequent edit and changed reliance unobserved.
• Evidence limits: No lint transcript, result, accuracy evaluation or downstream consumer trace.

Behavioral-authority path: external agent/user consumes the response; channel is chat proposal; force advisory; horizon current maintenance request, potentially later file edits. No automatic truth-admission license follows.

• /memwiki-lint: Perform a health check on the wiki. Look for outdated information, missing context in hot.md, or empty domain pages, and propose fixes. --- AGENTS.md

EPI-RTE-log-fold

Label: Work log condensation Part of: RT-RTE-agent-commands

This is the older-log transformation part; unlike lint it requests a changed retained object. Endpoints/progression: user token → model selects older log entries → summary replaces detail → possible later reading. Owner/executor: host agent and its external tools. Context/state: log and host task context. Candidate target: historical assertions. Evaluator/condition: model decides crucial history and condensation; no prescribed fidelity comparison or acceptance criterion. Semantic preservation is indeterminate without paired input/output. User supplies launch, model proposes/admits text through file tools; human veto remains host-dependent. Mode: open maintenance request; improvement trigger is excess log length, without a threshold. Guidance: preserve crucial history. What persists is summarized history; no required backup, verification or rollback. SRC-1, AGENTS.md, .memory/wiki/log.md.

• implementation conclusion status: claimed
• operation conclusion status: uninspected
• Immediate return: Requested summarized paragraph replacing older log detail.
• Later read-back: Summary is intended to remain in log.md for subsequent history access; actual read-back unobserved.
• Delegated visibility: uninspected — no delegated worker or sharing protocol is specified.
• Selection predicate: /memwiki-fold, then model selection of older entries and crucial history.
• Invalidation or expiry: Older detail may be replaced; no explicit provenance or reversible expiry metadata.
• Activation or effect: Requested change to durable history; resulting behavior dependence unobserved.
• Evidence limits: No before/after log, source-linked summary or fidelity check supplied.

Behavioral-authority path: later coding agent, retained log read through conditional wiki access, contextual/advisory force, subsequent project sessions. This is retention before acceptance, not post-acceptance integration.

• /memwiki-fold: Condense older entries in log.md into a summarized paragraph to keep the file from becoming too long, while preserving crucial history. --- AGENTS.md

Claims

EPI-CLM-project-memory

Label: Persistent project memory

Claimed operation: persistent, compounding project knowledge across coding sessions. Distinct from RT-CLM-safe-update, which concerns file overwrite safety, not epistemic growth. SRC-1, README.md, AGENTS.md; doctrine/design. The metaphor is operationally interpreted as persistence and later reuse, not an assertion that the CLI itself contains model intelligence.

memwiki gives your project a persistent brain. It leverages the fact that modern AI agents naturally read markdown files in your workspace, providing a standardized, vendor-agnostic memory system without requiring complex plugins or subscriptions. --- README.md

This project uses MemWiki (inspired by the LLM Wiki pattern), a persistent compounding knowledge base for AI coding agents. --- AGENTS.md

EPI-CLM-primary-truth

Label: Primary project truth

Claimed warrant/operation: protocol and hot cache receive primary-source priority; raw originals are called a source of truth. Distinct from RT-CLM-safe-update; closely associated with the reliance channel RT-BAP-host-protocol, but this record identifies its warrant language rather than another delivery path. SRC-1, CLAUDE.md, README.md; doctrine/design. Raw-file immutability does not independently validate content, and the protocol supplies no required claim-to-source citation.

Please read AGENTS.md and .memory/wiki/hot.md as your primary source of truth for this project before answering any queries or making changes. --- CLAUDE.md

• How it works: When you drop a file like stripe-api.pdf into .raw/, you simply tell your agent: "Please process the new Stripe docs in the raw folder." The agent will read the PDF and synthesize the important architectural details into a clean markdown file inside the wiki/ folder, leaving the original PDF untouched as a source of truth. --- README.md

EPI-CLM-health-fixes

Label: Wiki health fixes

Claimed operation: agent fixes stale information and missing context on lint. No supplied claim counterpart; RT-RTE-agent-commands already notes the stronger README wording. This is the distinct public claim compared against EPI-RTE-wiki-lint, not a duplicate command route. SRC-1, README.md, AGENTS.md; doctrine/design.

• /memwiki-lint: The agent will perform a health check on the wiki, fixing outdated information or filling in missing context. --- README.md

Evidenced absences

EPI-ABS-acceptance-gate

Label: Explicit acceptance gate

• implementation conclusion status: absent

Searched boundary: SRC-1 at 8034a3da991ac2639b87875172a9572903ecf1d5, complete reads of src/index.ts, src/index.js, AGENTS.md, README.md and all three host hooks. Query/question: does any candidate assessment yield a recorded, evidence-consuming acceptance decision naming criterion, intended use and scope, consumed before reliance; or a mandatory rejection/withdrawal marker? Dispatch and write branches use command/path/existence predicates only. Protocol requests synthesis, reads, health proposals and condensation, with no acceptance record or mandatory validity gate. Strongest counterexample is lint, retained in EPI-RTE-wiki-lint: it proposes fixes, rather than deciding explicit acceptance or preventing stale reuse. The closest supplied records RT-RTE-agent-session and RT-RTE-agent-commands describe open model admission, not an acceptance record. This absence is a distinct assessment of those mechanisms, not a replacement route.

Conclusion: no explicit acceptance gate in these shipped paths; prevents promotion of retention/use to accepted knowledge or post-acceptance integration. It does not establish that external hosts never evaluate or that all wiki assertions are false. Load-bearing branches and requested command passages are supplied by RT-RTE-cli-init, RT-RTE-protocol-update, RT-RTE-agent-commands, and the lint/fold records above.

• Synthesize any raw knowledge from .memory/.raw/ into the wiki pages.
• Never delete knowledge from the wiki. Only append or refine. --- AGENTS.md