Nuggets epistemic report

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

Source-and-claim boundary

The question is whether the shipped memory and gateway acquire, produce, check and license claims for later reliance. Scope and exclusions: see boundary; evidence identities and layers: see its Source register. All additional source inspection below is SRC-1, implementation unless labelled doctrine/design, at the same frozen commit. No execution or causal evidence was supplied or generated.

The assessment includes explicit memory operations, automatic capture, numerical reconstruction/cleanup, recall selection, withdrawal, migration, promotion, testing and gateway consumption. The absent advertised plugin, external Pi/provider internals, external Claude memory reader, live facts and execution traces prevent conclusions respectively about plugin behavior, model-internal checking, promoted-memory activation, actual candidate provenance and exercised benefit. Provider-generated propositions are possible, but their actual truth-apt content, derivations and checks are not determinable from an uninstantiated endpoint. Excluded publication articles remain outside this assessment. These are not system-complete negatives about every possible host behavior.

Consequential source claims are EPI-CLM-associative-recall and EPI-CLM-memory-improvement. Transport admission has a narrower supplied claim: see RT-CLM-ingress-control.

Entry point or operation	Source path	Coverage and records / limit
Root dispatch; advertised CLI	package.json	Direct read: scripts dispatch gateway, setup, migration and Vitest; no CLI bin declared. Boundary supplies tracked-tree evidence that documented Python/CLI implementation is not shipped. No inference from CLI doctrine to an executable evaluator.
Library remember, load, clear, forget, recall and shelf removal	src/nuggets/memory.ts, src/nuggets/shelf.ts	Direct reads of remember/recall/forget, serialization/rebuild/decode/key matching and full shelf; clear supplied by runtime/boundary and test assertions. EPI-RTE-fact-upsert, EPI-RTE-recall-decode, EPI-RTE-recall-score.
Numerical cleanup and softmax	src/nuggets/core.ts	Direct read of seeds, sharpening, CORVACS-lite, similarity normalization, softmax and bind/unbind; EPI-RTE-recall-score. Tests of other arithmetic supplied by boundary, not independently assessed.
All registered memory tool actions and lifecycle/input/file/compaction hooks	.pi/extensions/nuggets.ts	Direct reads of helper wrappers, extraction/rendering, reconstruction, all action dispatch and material hooks. UI rendering and overlay registration supplied by runtime, not a separate truth check. RT-RTE-memory-context, EPI-RTE-preference-capture, EPI-RTE-fact-withdrawal.
Kind classification and migration	src/nuggets/kinds.ts, src/migrate-memory.ts	Direct full reads; EPI-RTE-kind-migration.
Compaction promotion	src/nuggets/promote.ts	Direct full read and extension caller; RT-RTE-memory-promotion.
Schedule create/delete/list, prompt supply and confirmation	.pi/extensions/proactive.ts	Direct full read; RT-RTE-schedule-admission. Gateway admission and removal supplied runtime findings.
Message output and heartbeat filtering	src/gateway/router.ts	Direct full read; RT-RTE-message-turn, RT-RTE-proactive-turn. Other gateway files and mirrored gateway/src/ covered by supplied runtime, not fresh direct reads.
Setup and transport admission	src/setup.ts, src/gateway/config.ts	Supplied runtime findings: configuration/format/allowlist checks, not claim-truth evaluation. No credential-validity finding.
Memory regression tests	tests/memory.test.ts	Direct full read, not run; EPI-RTE-memory-tests. Core tests covered by supplied boundary only; no fresh assessment of their assertion completeness.
Public claims and instructions	README.md, NUGGETS_INSTRUCTIONS.md, CLAUDE.md	README claim passages directly read; remaining instruction claims supplied by boundary/runtime. Doctrine/design, not proof of operation.

Epistemic-object inventory

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source ID and local anchor	Gap/limit
EPI-OBJ-remember-value	Key/value may assert a scoped fact; arbitrary strings need not do so	see record	SRC-1 src/nuggets/memory.ts	No live item or evidence provenance supplied.
EPI-OBJ-preference-value	A preference attribution may be truth-apt; requested behavior itself is policy	see record	SRC-1 .pi/extensions/nuggets.ts	Regex does not establish speaker endorsement or preserve surrounding qualification.
EPI-OBJ-file-path	Path associated with a read/edit/write event; successful edit is not thereby established	see record	SRC-1 .pi/extensions/nuggets.ts	Input fields, not checked effects.
EPI-OBJ-task-summary	Truncated user text, potentially claims or requests; no actual proposition individuated	see record	SRC-1 .pi/extensions/nuggets.ts	Not injected through ordinary memory rendering.
RT-CMP-hrr-vectors	Numerical encoding of the stored values, not an independently warranted proposition	see record	SRC-1 src/nuggets/memory.ts	Decoding is associative selection, not factual confirmation.
EPI-OBJ-recall-score	Numerical score about relative decode similarity, not probability of world truth	see record	SRC-1 src/nuggets/memory.ts	No calibration or outcome reference.
EPI-OBJ-promoted-entry	Copies a retained value, with the same unresolved truth scope	see record	SRC-1 src/nuggets/promote.ts	Parseable old entries survive; unsupported prose may not. External reader uninspected.
RT-OBJ-cron-jobs	Symbolic scheduling fields and retained prompt, which may contain claims or requests	see record	SRC-1 .pi/extensions/proactive.ts	Scheduling criteria do not warrant prompt content.
RT-CMP-pi-model	Actual generated truth-apt content not determinable	see record	SRC-1 src/gateway/pi-rpc.ts	External distributed-parametric state; no pinned model or run evidence.

Authority-route ledger

Fields owned by the linked declaration are given there; the ledger separates consequential functions even when they share an implementation. All implemented routes below are unobserved in this run. “No epistemic license” means the named route supplies no evidence-consuming truth acceptance, not that every imported statement is false.

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
EPI-RTE-fact-upsert	content transformation	implemented	EPI-OBJ-remember-value	truth-apt transformation: acquisition/import	Caller value enters store	Trimming/nonempty and case-insensitive key rules; string domain	see record	New value or replacement	Enforcing write rule	Imported warrant unknown; no source check	Permits later use as a fact	see RT-BAP-facts-prompt	SRC-1 src/nuggets/memory.ts	none	none	Model authorship/provenance outside store.
EPI-RTE-fact-upsert	retention	implemented	EPI-OBJ-remember-value	no content change	In-memory facts to JSON	autoSave or explicit save	see record	Durable strings/configuration	Filesystem replacement	Retention, not acceptance	Makes values reloadable	Library/extension; JSON load; permissive; later invocations	SRC-1 src/nuggets/memory.ts	none	none	No evidence links or endorsement added.
EPI-RTE-preference-capture	content transformation	implemented	EPI-OBJ-preference-value	truth-apt transformation: indeterminate	Input fragment becomes preference/learning	First regex match, truncation and key generation	see record	Selected fragment and generated key	Enforcing extraction	General preservation not established; specific negated context can be lost	Admits text without per-item review	see RT-BAP-facts-prompt	SRC-1 .pi/extensions/nuggets.ts	none	Semantic-loss witness	Cannot call every extraction a faithful acquisition or a conjecture.
RT-RTE-memory-context	retention	implemented	EPI-OBJ-file-path	truth-apt transformation: acquisition/import	Tool path input to fact	Named tool and string path, not success status	see record	Basename-keyed path	Enforcing capture	Warrant degraded if read as successful work	Makes path available as Active Files	see RT-BAP-facts-prompt	SRC-1 .pi/extensions/nuggets.ts	none	none	Basename collisions overwrite; no effect oracle.
RT-RTE-memory-context	content transformation	implemented	EPI-OBJ-task-summary	truth-apt transformation: indeterminate	Last three selected user texts to 200 characters	Fixed slicing; no semantic evaluator	Before host compaction; see record	_task plus selected local file paths	Enforcing retention/selection	No semantic-summary license	Saves selected text; host compaction remains external	Shelf later recall/list; permissive; later invocations; ordinary prompt excludes _task	SRC-1 .pi/extensions/nuggets.ts	none	none	Omitted qualifications possible; not a model-generated summary.
EPI-RTE-recall-decode	content transformation	implemented	RT-CMP-hrr-vectors	truth-apt transformation: non-ampliative reshaping	Facts to numeric vocabulary/banks and selected value	Fixed seeded encoding, unbinding, numerical cleanup	see record	Existing vocabulary string or unresolved query	Numerical transformation	No novel factual entailment or truth check	Produces candidate recall result	Library/extension; recall result; permissive; current request	SRC-1 src/nuggets/memory.ts	EPI-CLM-associative-recall	none	Selection may associate wrong stored value; source truth unchanged.
EPI-RTE-recall-score	check/evidence production	implemented	EPI-OBJ-recall-score	no content change	Query/key match and value-vector similarity	Exact/substring/token/fuzzy selector; cosine/softmax over finite stored vocabulary	see record	found flag, confidence and margin	Ranking and match gate	Only within matching/decoding domain, not truth calibration	Enables result delivery and cross-store ranking	see record	SRC-1 src/nuggets/memory.ts	EPI-CLM-associative-recall	Confidence is not warrant	No accuracy threshold on returned softmax score.
RT-RTE-memory-context	operational admission/selection/consumption	implemented	EPI-OBJ-remember-value	no content change	Shelf result or local substring fallback; prompt formatting	Highest confidence across ordered kinds; ties retain earlier result; local fallback only if shelf fails	see record	Answer with confidence/source label or selected facts in system prompt	Ranking/permissive context	No evidence-consuming acceptance	Supplies retained text to model before another source check	see RT-BAP-facts-prompt	SRC-1 src/nuggets/shelf.ts	none	Priority is not first-kind-only	Model receives value and coarse kind/name, not original evidence lineage.
EPI-RTE-fact-withdrawal	disposition/acceptance	implemented	EPI-OBJ-remember-value	no content change	Key deletion, not truth acceptance	Model-requested forget; local delete OR shelf delete	see record	Local deletion can skip durable deletion	Enforcing locally, incomplete durable withdrawal	No falsity judgment or acceptance	Can remove current injection while retained copy remains	see record	SRC-1 .pi/extensions/nuggets.ts	none	Retraction gap	History reconstruction and promotion can retain older copies.
EPI-RTE-kind-migration	lineage/freshness/recovery	implemented	EPI-OBJ-remember-value	truth-apt transformation: non-ampliative reshaping	Legacy file to kind stores	Prefix/value heuristic and case-insensitive merge	see record	Buckets; incoming value wins conflict; backup/archive	Enforcing file update	No additional warrant	Changes which default-kind consumer finds an entry	see record	SRC-1 src/migrate-memory.ts	none	none	Preserves/max-merges counters, not evidential provenance.
RT-RTE-memory-promotion	check/evidence production	implemented	EPI-OBJ-promoted-entry	no content change	Entry hit count at least three	Deterministic usage threshold, not truth criterion	see record	Candidate qualifies or not	Selection gate	Reuse count only	Permits promotion writer	Writer; hits field; enforcing; compaction event	SRC-1 src/nuggets/promote.ts	EPI-CLM-memory-improvement	Ordinary recall omits session ID	Distinct-session uniqueness is not globally tracked.
RT-RTE-memory-promotion	retention	implemented	EPI-OBJ-promoted-entry	truth-apt transformation: non-ampliative reshaping	Selected entries to Markdown	Existing recognized sections/entries merged with candidates	see record	Rewritten MEMORY.md	Conditional file write	Not lifecycle integration after acceptance	Offers text for intended external context	External Claude reader; Markdown; intended advisory force; later sessions unestablished	SRC-1 src/nuggets/promote.ts	EPI-CLM-memory-improvement	Host consumption uninspected	Old promoted values are not removed when source is absent; free-form prose can disappear.
EPI-RTE-memory-tests	check/evidence production	implemented	EPI-OBJ-remember-value	no content change	Fixture-defined memory behavior	Vitest assertions against explicit expected outputs	see record	Would pass/fail assertions	Test-run failure only	Fixture behavior domain, not world truth	No shipped fact-admission consumer	Developer; test result; advisory to development; no runtime acceptance path	SRC-1 tests/memory.test.ts	none	none	Not executed; no causal benefit finding.
RT-RTE-schedule-admission	operational admission/selection/consumption	implemented	RT-OBJ-cron-jobs	non-truth-apt policy/content update: scheduled future request	Append success versus job admission	Extension shell exit versus gateway fields/default JID	see record	Confirmation before acceptance; possible later rejection	Local operational gate	Does not warrant confirmation of actual scheduled delivery	Can create/delete retained prompts for later Pi turns	see RT-BAP-scheduled-prompt	SRC-1 .pi/extensions/proactive.ts	none	Confirmation exceeds checked scope	Intended time, recipient and completion not checked by an oracle.
RT-RTE-proactive-turn	operational admission/selection/consumption	implemented	RT-OBJ-cron-jobs	no content change	Model response admitted to outgoing chat	For heartbeat: nonempty and does not contain NOTHING, case-insensitive	see record	Send or suppression	Enforcing output filter	No relevance/truth acceptance	Allows or suppresses delivery	Router; response text; enforcing; current proactive turn	SRC-1 src/gateway/router.ts	none	Substring sentinel	Nothing failed; deployment succeeded is suppressed by deduction, not observed execution.

No lifecycle integration after evidence-consuming truth acceptance was found in the inspected local memory routes: see EPI-ABS-truth-acceptance. This is distinct from implemented retention, retrieval and operational use.

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-associative-recall	see record	SRC-1 README.md; doctrine/design	see record	EPI-RTE-recall-decode, EPI-RTE-recall-score	None supplied	No benchmark, intervention or calibrated error data	Algebraic retrieval from finite stored vocabulary is implemented; tests assert selected examples	Speed/accuracy not established. Stored fact truth is outside numerical cleanup's domain.
EPI-CLM-memory-improvement	see record	SRC-1 README.md; doctrine/design	see record	RT-RTE-memory-promotion, RT-RTE-memory-context	None supplied	No later behavior, cost/latency baseline or component variation; external reader missing	A conditional hit-selected context writer and independent per-turn memory injection exist	Ordinary extension recall does not accrue hits. Permanent every-session context and faster/cheaper behavior exceed evidence.

Bounded conclusion

Nuggets acquires caller strings and user/tool fragments, persists symbolic records and deterministically reconstructs numerical HRR state. Recall selects among stored strings; it is neither a source-truth oracle nor a route to new entailed world claims. The confidence field measures relative decoding scores. Numeric cleanup changes signals, not evidential support. See EPI-RTE-fact-upsert, EPI-RTE-recall-decode and EPI-RTE-recall-score.

The routes can carry truth-apt facts, preference attributions and file-context claims, but they do not retain claim-specific source evidence. Explicit imports have unknown warrant. Automatic extraction can degrade it: surrounding negation and attribution are not checked, file captures need not reflect successful work, and the task excerpt truncates without a semantic check. Actual model-generated conjectures and provider-internal evaluation remain not determinable; source text permits no attribution of a particular candidate to that opaque processing. See EPI-RTE-preference-capture and the object inventory.

Operational reliance is concrete and separate from epistemic acceptance. Prompt injection supplies selected facts unsolicited on later turns; requested recall delivers selected values with a coarse source label. User preferences can directly shape requested policy without a truth-apt transformation. Neither route enforces compliance or demonstrates improved future action. Key replacement, migration and promotion change retained context without evidence-consuming truth acceptance. Forget is not a reliable cross-copy retraction. See RT-BAP-facts-prompt, EPI-RTE-kind-migration, EPI-RTE-fact-withdrawal and EPI-ABS-truth-acceptance.

Three hits are an operational selection criterion for intended permanent context, not a criterion accepting a proposition as true. The writer retains selected values and parseable old entries; the supplied boundary does not include actual Claude read-back. Even hit counting tracks only the last session ID, not the set of sessions; alternating two IDs can repeatedly increment it. Replacing a value does not reset its counter. These counterexamples limit the claim independently of ordinary extension calls' missing session ID. See RT-RTE-memory-promotion, EPI-RTE-fact-upsert and EPI-CLM-memory-improvement.

Scheduling and output filters have consequential but non-epistemic authority. Append success licenses a creation confirmation before gateway admission, and heartbeat substring filtering can suppress meaningful text without judging its truth or relevance. Test source supplies expected-answer oracles for fixtures, but no runtime fact-admission consumer or executed success evidence. See RT-RTE-schedule-admission, RT-RTE-proactive-turn and EPI-RTE-memory-tests.

For the README's self-improvement claim, the assessment boundary is the shipped context-selection machinery across future sessions; its independently specifiable objective is lower time and cost for useful action. Usage evidence can select a context write, but usefulness is not evaluated; later host reading and behavior causally depending on that write are unestablished. This supports a conditional improvement-directed context pathway, not exercised or successful self-improvement. Learning benefit, content-directed criticism, reflection and autonomous theory-building do not follow from persistence or autonomous writes. Provider cognition remains uninspected rather than absent. These independent limits do not negate the implemented memory and policy-supply routes.

Shared records

Operative objects

EPI-OBJ-remember-value

Label: Remember value

Part of: RT-RTE-memory-context

This is the explicit caller-authored value handled inside that integration route (object within a route, not a duplicate route). Natural-language values can carry propositions or instructions; keys are symbolic routing/matching units. Storage is the local Map, tool-result snapshots and shelf JSON facts. Content truth depends on the actual key/value and scope; none was supplied. SRC-1 src/nuggets/memory.ts, .pi/extensions/nuggets.ts; load-bearing admission passage in EPI-RTE-fact-upsert.

EPI-OBJ-preference-value

Label: Preference value

Part of: RT-RTE-memory-context

A distinct automatically extracted natural-language fragment and generated symbolic preference/learning key. It is stored in local facts and user-kind JSON, rendered as Preferences or Learnings, and interpreted by Pi/model. It differs from the explicit value by producer and extraction checks. SRC-1 .pi/extensions/nuggets.ts; passages in EPI-RTE-preference-capture.

EPI-OBJ-file-path

Label: File path

Part of: RT-RTE-memory-context

Symbolic path input captured under basename-derived keys, persisted by tool-result hooks to project memory and held locally; compaction also copies selected calls' paths locally. Paths are fixed string inputs to capture, then natural-language Active Files context to the model. This records association with attempted operations, not checked filesystem effects. SRC-1 .pi/extensions/nuggets.ts.

if (event.toolName === "read") { facts.set(file:${basename}, filePath); shelfRemember(file:${basename}, filePath, "project"); } else if (event.toolName === "edit" || event.toolName === "write") { facts.set(edited:${basename}, filePath); shelfRemember(edited:${basename}, filePath, "project"); } --- .pi/extensions/nuggets.ts

EPI-OBJ-task-summary

Label: Task summary

Part of: RT-RTE-memory-context

Natural-language excerpt stored under symbolic _task in local facts and project JSON before host compaction. It is not an inferred summary; it may contain requests rather than claims. Ordinary injection excludes underscore-prefixed entries. Recall or explicit listing can still expose it. SRC-1 .pi/extensions/nuggets.ts.

const taskSummary = userMessages.slice(-3).join(" | ").slice(0, 200); facts.set("_task", taskSummary); shelfRemember("_task", taskSummary, "project"); --- .pi/extensions/nuggets.ts

EPI-OBJ-recall-score

Label: Recall score

A symbolic numeric result (confidence, margin, found) held transiently in return objects and formatted into tool text. It is a distinct part of the memory integration's recall result, not the retained value or RT-CMP-hrr-vectors. Scope includes direct library callers, so it is not wholly contained in the extension integration. SRC-1 src/nuggets/memory.ts; passages in EPI-RTE-recall-score.

EPI-OBJ-promoted-entry

Label: Promoted entry

Part of: RT-RTE-memory-promotion

Natural-language value under a section/key in MEMORY.md, with Markdown providing localized organization. This object is the output part of the supplied route, not a second declaration of that route. Intended consumer is external Claude; actual loading is not inspected. SRC-1 src/nuggets/promote.ts.

for (const [key, value] of entries) { lines.push(- **${key}**: ${value}); } --- src/nuggets/promote.ts

Routes

EPI-RTE-fact-upsert

Label: Fact upsert

The library route accepts caller key/value → trimmed fact list → optional JSON save. It overlaps the explicit-write step of RT-RTE-memory-context, but also admits direct library writes, so it is not the same referent or wholly contained there. Owner is deterministic library code; caller/model proposes content. It checks empty strings, not truth, with no per-content human veto or answer oracle. Mode is ordinary open use; replacement and optional oldest-entry eviction are admission policy. Caller may overwrite/forget; no evidential rollback decision exists. Extension guidance asks for useful discoveries and short values, not evidence-bearing acceptance.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: remember returns void; extension separately reports remembered content.
• Later read-back: Saved JSON reloads values for reconstruction, recall and prompt supply.
• Delegated visibility: Library receives strings, not underlying source evidence; extension/model later receives selected text.
• Selection predicate: Trimmed nonempty fields; case-insensitive key update; optional maximum-facts tail.
• Invalidation or expiry: Value replacement, forget/clear or optional oldest-entry eviction; updates retain prior hit metadata.
• Activation or effect: Facts and optionally disk content change; no observed downstream behavior.
• Evidence limits: No live content, provenance or success traces. Classification is import at the store edge; caller production may be model-generated but is not inspected here.

SRC-1 src/nuggets/memory.ts, .pi/extensions/nuggets.ts.

key = key.trim(); value = value.trim(); if (!key || !value) return; --- src/nuggets/memory.ts

if (f.key.toLowerCase() === key.toLowerCase()) { f.value = value; found = true; break; } --- src/nuggets/memory.ts

max_facts: this.maxFacts, facts: this._facts, config: { --- src/nuggets/memory.ts

EPI-RTE-preference-capture

Label: Preference capture

Part of: RT-RTE-memory-context

Input text → first matching preference/learning expression → generated key/truncated value → local and user shelf writes. Owner is extension code, not an explicit model judgment. Non-extension input triggers automatic admission, without per-item human approval, oracle or semantic veto. Mode is ordinary use; source patterns supply all guidance. Overwrite/forget can change retained material, subject to the withdrawal limits below.

Concrete static trace: the input Do not remember that deployment is safe matches the unanchored remember that pattern. Captured value is deployment is safe; short key is deployment-is-safe; retained key is learn:deployment-is-safe. Rendering places that value under Learnings, without the negated instruction. The input does not entail endorsement of deployment safety. This is an implementation deduction, not a target execution or evidence that such an item exists. General extraction is classified indeterminate: an unqualified actual preference can be faithfully imported, while negation, quotation and truncation can change apparent meaning. Actual input/context is required to decide each edge.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Input hook continues the event after automatic local/shelf update.
• Later read-back: User-kind facts hydrate local state and reach future prompt rendering or recall.
• Delegated visibility: Later model receives generated preference/learning key and value, not the original surrounding input.
• Selection predicate: First matching regex; captured value trimmed and limited to 100 characters; key derived from first 30.
• Invalidation or expiry: Same-key overwrite or deletion; no stored link allowing automatic withdrawal when original context changes.
• Activation or effect: Automatic retention and potential prompt supply are wired; actual model reliance is unobserved.
• Evidence limits: Deterministic counterexample refutes universal semantic preservation; it does not establish every extraction is ampliative.

SRC-1 .pi/extensions/nuggets.ts.

{ regex: /\bremember that (.+)/i, prefix: "learn" }, --- .pi/extensions/nuggets.ts

const value = match[1].trim().slice(0, 100); const shortKey = value.slice(0, 30).replace(/\s+/g, "-").toLowerCase(); return { key: ${prefix}:${shortKey}, value }; --- .pi/extensions/nuggets.ts

const pref = extractPreference(event.text); if (pref) { facts.set(pref.key, pref.value); shelfRemember(pref.key, pref.value, "user"); } --- .pi/extensions/nuggets.ts

EPI-RTE-recall-decode

Label: Recall decode

Library facts → unique value vocabulary/seeded vectors → superposed banks → unbinding/cleanup → best vocabulary string. Fixed code owns the route. Distinct from the supplied RT-CMP-hrr-vectors, which identifies its numerical state, and broader than the extension-only integration in RT-RTE-memory-context. Values update the representation during operation, but there is no provider-weight training here. Algorithm is frozen; exact live vectors depend on unsupplied facts/configuration. No mutable model endpoint is used by this route.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Selected existing vocabulary word, score/margin and matched key, or empty result.
• Later read-back: Facts/configuration are persisted; vectors are reconstructed rather than serialized. Returned answer is not automatically saved as a novel fact.
• Delegated visibility: Caller receives a string/result, not the full vector banks or source evidence.
• Selection predicate: Key resolution followed by maximum softmax score among stored unique values.
• Invalidation or expiry: Dirty state triggers rebuild after mutations; removing all facts clears numerical state.
• Activation or effect: Produces a recall candidate; extension may deliver it as tool text.
• Evidence limits: Selection is non-ampliative at the vocabulary edge but can misassociate a key/value. No truth check, benchmark or observed accuracy follows.

SRC-1 src/nuggets/memory.ts, src/nuggets/core.ts.

// Build vocabulary from unique values const seen = new Set(); const vocab: string[] = []; for (const f of this._facts) { if (!seen.has(f.value)) { vocab.push(f.value); seen.add(f.value); } } this._vocabWords = vocab; --- src/nuggets/memory.ts

return { word: this._vocabWords[bestIdx], sims: simsSum, probs }; --- src/nuggets/memory.ts

EPI-RTE-recall-score

Label: Recall score

Query/keys and decoded numerical vector → match decision plus similarity-derived score → shelf ranking and tool display. This is the evaluator part of recall, distinct from EPI-RTE-recall-decode's string-producing transformation; both overlap RT-RTE-memory-context but include direct library consumers. Target/domain: query-to-key lexical resemblance and reconstructed signal-to-vocabulary similarity, not a proposition about the world. Program rules perform evaluation; no model/human truth judgment, reference world answer or per-item veto occurs. Ordinary open recall is the operating mode.

A static single-value case exposes the license boundary: a sole stored value is the only vocabulary candidate, so softmax has no competing factual alternative. Its near-one normalized score cannot establish that value is true. Key matching also has its own gate; a high score cannot rescue an unresolved key. Conversely a resolved key returns found: true without requiring a minimum confidence or margin. Across kinds, recallInOrder retains the greatest confidence, not simply the first matching kind. Earlier kind order matters on ties.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Found flag, confidence from top probability, and top-two margin.
• Later read-back: Scores are not retained as evidence; with a nonempty session ID the matched key's usage counter is saved for possible promotion.
• Delegated visibility: Model sees answer and rounded confidence/coarse source label; margin is not included in that extension text.
• Selection predicate: Exact/substring then token overlap at least 0.5 then fuzzy threshold; within values use maximum softmax; across shelves use greatest confidence.
• Invalidation or expiry: Recomputed from current facts/configuration; counters remain on value replacement and deduplicate only against the last session ID.
• Activation or effect: Match gates and confidence ranking can change returned content; usage counters can qualify entries for promotion.
• Evidence limits: No calibration, world-truth criterion or executed accuracy/benefit comparison. Numerical cleanup's target is signal magnitude; default sharpening and CORVACS-lite can be identity functions.

SRC-1 src/nuggets/memory.ts, src/nuggets/shelf.ts, src/nuggets/core.ts, .pi/extensions/nuggets.ts.

const probs = softmaxTemp(simsSum, this._tempT); let bestIdx = 0; for (let i = 1; i < V; i++) { if (probs[i] > probs[bestIdx]) bestIdx = i; } --- src/nuggets/memory.ts

return { answer: word, confidence, margin, found: true, key: tag }; --- src/nuggets/memory.ts

if (!best.found || result.confidence > best.confidence) { best = { ...result, nugget_name: name }; } --- src/nuggets/shelf.ts

if (p === 1.0) return z; --- src/nuggets/core.ts

if (a <= 0) return z; --- src/nuggets/core.ts

EPI-RTE-fact-withdrawal

Label: Fact withdrawal

Part of: RT-RTE-memory-context

Explicit forget request → key requirement → short-circuit local/shelf deletion → report. This disposition part does not judge falsity. Model/caller proposes key, extension chooses deletion path, with no separate human veto or oracle. Mode is ordinary open use. Recovery is later overwrite/reload, not an atomic cross-copy withdrawal. Supplied runtime already identifies the short-circuit; additional direct inspection shows reconstruction merges positive snapshots without deletion tombstones.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Reports Forgot if either attempted branch returns true; details contain only remaining current local facts.
• Later read-back: Session start can hydrate the undeleted shelf value; branch reconstruction can retain earlier snapshot keys absent from later snapshots.
• Delegated visibility: Model sees success text and remaining snapshot, not a cross-store withdrawal receipt.
• Selection predicate: Required key; local exact-key delete first, shelf helper only on false.
• Invalidation or expiry: No cascading removal of promoted Markdown; shelf deletion may remove only the first matching kind; history has no explicit tombstone for normal forget.
• Activation or effect: Local prompt content can change immediately while future shelf/history read-back revives it.
• Evidence limits: Static reachability, not an observed reappearance. Library direct forget does remove matching case-insensitive facts, so not every deletion fails.

SRC-1 .pi/extensions/nuggets.ts, src/nuggets/memory.ts; short-circuit passage supplied in RT-RTE-memory-context.

for (const [k, v] of Object.entries(details.facts)) { if (v) { facts.set(k, v); } else { facts.delete(k); } } --- .pi/extensions/nuggets.ts

EPI-RTE-kind-migration

Label: Kind migration

Operator-launched legacy JSON → backup → heuristic buckets → case-insensitive merge into kind JSON → archive. Distinct from the supplied extension/promotion routes; runtime mentions this driver without declaring a matching route. Code owns classification/admission, operator starts the operation but does not approve individual facts. Guidance is key prefixes/value slash matching; no truth oracle or content criticism. Ordinary maintenance, not a bounded improvement experiment. Backups permit manual recovery; no automatic rollback on partial write is implemented.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Console paths/counts, or an early no-file return.
• Later read-back: Default-kind stores are loaded by the shelf and consumed by extension recall/hydration.
• Delegated visibility: Later reader sees the winning value and merged usage metadata, not the displaced alternative or merge rationale.
• Selection predicate: inferMemoryKind; same lowercased key merges, incoming value wins, hits become maximum.
• Invalidation or expiry: Legacy file archived after writes; backup retained. Subsequent writes can replace resulting values.
• Activation or effect: Reorganizes retained claims and can resolve conflicts by overwrite without evaluating their truth.
• Evidence limits: No live legacy file or migration trace; pure values are reshaped/selected, not accepted or semantically derived.

SRC-1 src/migrate-memory.ts, src/nuggets/kinds.ts.

target[idx] = { key: incoming.key, value: incoming.value, hits: Math.max(existing.hits ?? 0, incoming.hits ?? 0), last_hit_session: incoming.last_hit_session || existing.last_hit_session || "", }; --- src/migrate-memory.ts

EPI-RTE-memory-tests

Label: Memory tests

Developer-triggered Vitest → constructed memory fixtures → assertions. This evaluator route has no supplied canonical counterpart; boundary lists it as inspected test source, while runtime declares production routes. Test authors supply reference expected outputs, code judges results, developer can revise/reject changes; no automatic admission into the running memory store follows. Mode is bounded fixture tests, triggered by the package test script, not an online improvement loop. Target/domain: specified remember/recall/upsert/forget/save/load/hit-count/matching behavior for finite examples.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Would pass/fail assertions under Vitest; not executed here.
• Later read-back: inapplicable — test results are not wired to retained fact admission or later model prompts.
• Delegated visibility: Test runner receives fixture state and expected values; production model receives no test certificate.
• Selection predicate: Shipped test cases and explicit assertions, not random user claims or source evidence.
• Invalidation or expiry: Fixtures are recreated per test and temporary directories removed; no production rollback consumer.
• Activation or effect: Assertions can fail a developer test command, not block ordinary remember calls.
• Evidence limits: Source assertion existence does not establish passing execution, broad generalization or factual truth of arbitrary retained content.

SRC-1 tests/memory.test.ts, package.json.

n.remember("color", "blue"); const result = n.recall("color"); expect(result.found).toBe(true); expect(result.answer).toBe("blue"); expect(result.confidence).toBeGreaterThan(0); --- tests/memory.test.ts

Claims

EPI-CLM-associative-recall

Label: Associative recall

Claim: HRR stores individually retrievable facts and recall is algebraic/sub-millisecond. SRC-1 README.md, doctrine/design. Distinct from RT-CLM-ingress-control, whose subject is transport admission. Implementation supports finite-vocabulary associative recall, not a calibrated knowledge warrant or measured speed.

• Holographic memory — facts are stored as superposed complex-valued vectors using Holographic Reduced Representations (HRR). Recall is algebraic, sub-millisecond, and runs locally with zero external dependencies. --- README.md

EPI-CLM-memory-improvement

Label: Memory improvement

Claim: repeated recalls promote a fact to permanent context, making the agent faster and cheaper. SRC-1 README.md, doctrine/design. It overlaps the subject of RT-RTE-memory-promotion but is a claim about its effect, not that route's implementation identity; no matching supplied claim exists.

The writer selects three-hit entries and reconstructs recognized Markdown sections. This creates operational placement without truth acceptance. Additional static counterexamples: last_hit_session deduplicates only consecutive identical IDs, so alternating A/B/A can count three hits from two distinct sessions; upsert changes value without resetting that key's hits. Existing promoted keys remain when source facts disappear. Parsing accepts only section headers and formatted fact lines: given existing free prose plus a new qualifying candidate, the rewrite can omit the prose. No automatic withdrawal linkage exists.

• Self-improving — when a fact is recalled 3+ times across sessions, it gets promoted to permanent context. The agent gets faster and cheaper over time. --- README.md

if (fact.last_hit_session !== sessionId) { fact.hits = (fact.hits || 0) + 1; fact.last_hit_session = sessionId; if (this.autoSave) this.save(); } --- src/nuggets/memory.ts

const factMatch = stripped.match(/^-\s+**(.+?)**:\s*(.+)$/); if (factMatch && currentSection) { sections[currentSection][factMatch[1].trim()] = factMatch[2].trim(); } --- src/nuggets/promote.ts

Threshold and ordinary-call evidence: see RT-RTE-memory-promotion. Proposed self-change is context selection over future sessions; lower action time/cost is the separately stated objective. Use evidence can shape selection and a file can change, but usefulness assessment, external loading and later causal dependence are not established. No exercised self-improvement or benefit attribution follows.

Evidenced absences

EPI-ABS-truth-acceptance

Label: Truth acceptance

• implementation conclusion status: absent

Bounded absence: no evidence-consuming decision accepting a truth-apt value against a named truth/reliability criterion for a stated use/scope was found in the local memory implementation and extension routes inspected at SRC-1's frozen revision. Search boundary: src/nuggets/memory.ts, src/nuggets/shelf.ts, src/nuggets/kinds.ts, src/nuggets/promote.ts, src/migrate-memory.ts, material dispatch/hooks in .pi/extensions/nuggets.ts; test comparison in tests/memory.test.ts. Inspection question was whether any return/write/rejection branch consumes evidence about candidate truth rather than nonemptiness, key/string resemblance, signal similarity or usage count. The routes above enumerate those actual checks and consumers. Load-bearing passages: EPI-RTE-fact-upsert, EPI-RTE-recall-score and supplied RT-RTE-memory-promotion.

This is distinct from any supplied record: runtime declared no absence. It supports absence of local evidence-consuming truth acceptance and hence of lifecycle integration after such acceptance in these inspected paths. It does not deny that an external model/user may evaluate content before proposing a write, or that hosted tools may return evidence; those internal processes and actual cases are uninspected. It also does not deny operational admission, regression-test evaluation, policy supply or useful recall.