MemPalace epistemic report

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

Source-and-claim boundary

The question is what licenses reliance on retained conversation and project content, and what checks actually change its admission or later use. Scope and exclusions are those in boundary; sources are declared in its Source register. All added inspection uses SRC-1, implementation unless explicitly labelled doctrine/design. No target or evaluator was executed.

The supplied runtime is a provisional route account, not a coverage limit. Direct inspection of CLI and MCP fragment loading found additional material functions: checkpoint near-duplicate rejection, explicit drawer replacement/deletion, KG temporal withdrawal/replacement, and navigability audit with a failing exit threshold. These refine rather than supersede the runtime's valid containers. A claim that MemPalace merely appends or only returns text would miss these functions. A claim that it verifies stored propositions would confuse structural, similarity and temporal checks with evidence-consuming acceptance.

Public warrant claims include RT-CLM-recall-protocol and EPI-CLM-benchmark-recall. The capability claim RT-CLM-readonly-boundary limits operational admission, not epistemic endorsement.

Entry point or operation	Source path	Coverage and supporting records, or prevented conclusion
CLI dispatch	mempalace/cli/__init__.py	Direct read of fragment list; mining/setup/task routes supplied by runtime. Individual parser branches not fully inspected here.
Full/light MCP dispatch and registered handlers	mempalace/mcp_server/__init__.py, mempalace/mcp_light_server.py	Direct fragment list and result-enrichment read; supplied RT-RTE-mcp-dispatch. Registry includes KG and diary, not only drawer tools. Full schema catalog not independently read.
Turn, transcript and setup admission	mempalace/integrations/hermes/__init__.py, mempalace/hooks_cli.py, mempalace/cli/cmd_init.py	Supplied RT-RTE-hermes-capture, RT-RTE-hook-ingest, RT-RTE-init-detection; source excerpts supplied, not fresh full reads.
Drawer add/update/delete and structural checking	mempalace/mcp_server/tools_write.py, mempalace/config.py	Direct reads; EPI-RTE-drawer-revision, with RT-RTE-mcp-dispatch.
Semantic duplicate check and checkpoint skip	mempalace/mcp_server/tools_read.py, mempalace/mcp_server/tools_diary.py	Direct reads; separate EPI-RTE-duplicate-check and EPI-RTE-checkpoint-disposition. Other diary implementation remains unassessed.
Query/prefetch/wake-up and recall instructions	mempalace/searcher/query.py, mempalace/searcher/candidates.py, skills/mempalace-recall/SKILL.md	Supplied RT-RTE-palace-recall and RT-BAP-agent-context; direct light-wrapper read adds EPI-RTE-relevance-tags. No complete backend-specific recall assessment.
KG add/invalidate/supersede/query	mempalace/mcp_server/tools_kg.py, mempalace/knowledge_graph.py	Direct handler reads and add/invalidate implementation; supersede read through boundary validation, not its complete transaction or underlying query SQL. EPI-RTE-kg-temporal. Atomic/as-of guarantee remains partly uninspected.
Audit and tunnel maintenance	mempalace/cli/cmd_audit.py, mempalace/palace_audit.py, mempalace/cli/cmd_tunnels.py	Direct audit consumer and score function reads; EPI-RTE-palace-audit. EPI-RTE-tunnel-maintenance directly covers proposal review, --yes apply/prune and stale-wing exclusion; underlying proposal/prune algorithms not inspected, preventing validity or completeness conclusions for inferred associations.
Retrieval/model evaluation	benchmarks/longmemeval_bench.py, benchmarks/model_eval/runner.py, README.md	Direct selected metric/consumer reads; EPI-RTE-retrieval-evaluation, EPI-CLM-benchmark-recall. Model runner directly dispatches labelled calibration/classification/entity/memory scoring; its scorers and providers not inspected. No verified model-quality conclusion.
Coordination, patch delivery and backend capability admission	mempalace/tasks.py, mempalace/logstream.py, mempalace/backends/registry.py	Supplied RT-RTE-task-launch, RT-RTE-patch-delivery, RT-RTE-backend-registration; no new truth-acceptance finding.
Other included material families	mempalace/miner.py, mempalace/layers.py, crates/, mempalace/backends/, benchmarks/	Project/document/source-adapter transformations, closet/AAAK generation, native/media/code routes, non-Chroma variants, replication, repair/backup, other evaluators and host shell variants remain partly or wholly unassessed here. Available evidence is not an access gap. Their omission prevents a system-complete negative about derivation, conjecture, acceptance or evaluator routes.

Missing evidence is local to conclusions: no live corpus prevents assessment of the actual truth of retained claims; no host trace prevents demonstrated compliance or behavioral activation; no rerun/dataset inspection prevents reproduced recall or encoding-fidelity conclusions; no controlled intervention prevents attributing improvement to a component. Exact external model artifacts and provider internals are unavailable within this pin, preventing claims about their learning, knowledge or stable behavior. Distributed-parametric fixity is separately assessed in RT-CMP-minilm, RT-CMP-embeddinggemma, RT-CMP-embeddinggemma2, RT-CMP-openai-compat and RT-CMP-llm-provider. Fixed inference weights do not imply absence of behavior change through retained text.

Epistemic-object inventory

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source ID and local anchor	Gap/limit
EPI-OBJ-drawer-content	User/assistant testimony or caller-authored propositions, where the text states propositions	Recallable drawer	SRC-1 mempalace/mcp_server/tools_write.py	Particular truth-apt claims not determinable without corpus; attribution is not verification.
EPI-OBJ-kg-triple	Subject–predicate–object relationship over a caller-specified interval	Fact with optional provenance	SRC-1 mempalace/knowledge_graph.py	Default numerical confidence is not calibrated evidence; predicate interpretation and external truth unchecked.
EPI-OBJ-search-tags	Algorithm-relative relevance class and recency rank, not truth probability	Decorate recall result	SRC-1 mempalace/mcp_light_server.py	Distance and timestamps are the inputs; no endorsement or supersession follows.
EPI-OBJ-audit-report	Computed organization scores relative to fixed formulas	Navigability report	SRC-1 mempalace/palace_audit.py	Formula target differs from truth or actual answer quality.
EPI-OBJ-retrieval-metrics	Label-relative retrieval hit metrics	Benchmark evidence	SRC-1 benchmarks/longmemeval_bench.py	Label/source correctness, actual run and live-loop transfer not established.
RT-OBJ-task-request	Goal/done/delivery instructions; any embedded propositions depend on task body	Worker guidance	see RT-OBJ-task-request	Instruction force is not proposition acceptance.
RT-OBJ-patch-artifact	Diff is an executable proposal; accompanying rationale may be truth-apt	Delivered change proposal	see RT-OBJ-patch-artifact	Persistence does not establish correct code or installation.

Authority-route ledger

Rows separate checking, disposition and retention even when one handler connects them. Architectural status describes wiring, not observed activation. Repeated record-owned fields are supplied by the cited declaration; ledger-local judgments below state the epistemic license.

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	Behavioral-authority path	Evidence source ID and local anchor	Claim IDs	Mismatch marker	Gap/limit
RT-RTE-hermes-capture, RT-RTE-hook-ingest	content transformation	implemented	EPI-OBJ-drawer-content	truth-apt transformation: acquisition/import	Host/source content enters palace	see cited records	see cited records	see cited records	Filing request	Source warrant unknown; host labels preserve attribution, not truth	May retain user and model text without per-item approval	see cited records	see cited records	RT-CLM-recall-protocol	Lossless capture not guaranteed	No corpus/host trace
RT-RTE-mcp-dispatch	check/evidence production	implemented	EPI-OBJ-drawer-content	no content change	Input structure, duplicate ID, post-write readability	Software validates size/type/null bytes, identity and storage IDs; not truth	see record; add handler before/after upsert	Error, existing ID, readable ID	Handler refuses malformed/unreadable writes	Only structural validity or detected storage visibility	Blocks add or reports success/error	Caller through MCP result, enforcing current request	SRC-1 mempalace/config.py, mempalace/mcp_server/tools_write.py	none	Storage success is not acceptance	Backend atomicity remains an external contract
RT-RTE-mcp-dispatch	retention	implemented	EPI-OBJ-drawer-content	truth-apt transformation: non-ampliative reshaping	Validated drawer to bounded chunks	Deterministic slicing; see record for admission	see record	see record	Durable upsert	No added warrant; textual retention before acceptance	Later search/direct reads allowed	see RT-BAP-agent-context	SRC-1 mempalace/mcp_server/tools_write.py	RT-CLM-recall-protocol	none	Sanitization can alter invalid encoding; no universal fidelity claim
EPI-RTE-duplicate-check	check/evidence production	implemented	EPI-OBJ-drawer-content	no content change	Similarity of candidate to up to five retrieved drawers	see record	see record	see record	Diagnostic on direct call	Similarity relative to model/metric; not equivalence or truth	Supplies checkpoint's skip predicate	see record	see record	none	Semantic sameness unestablished	Contradictory or materially changed wording can be close in embedding space
EPI-RTE-checkpoint-disposition	disposition/acceptance	implemented	EPI-OBJ-drawer-content	no content change	Whether item is filed	see record	see record	see record	Enforcing skip or add call	Admission/rejection only; no acceptance against truth evidence	Suppresses duplicate-marked candidates; dedup error does not veto filing	see record	see record	RT-CLM-recall-protocol	Approximate match can suppress a correction	No preserved rejected-candidate text in returned duplicate entry
EPI-RTE-drawer-revision	lineage/freshness/recovery	implemented	EPI-OBJ-drawer-content	truth-apt transformation: acquisition/import	Caller replacement or deletion of retained text	see record	see record	see record	Enforcing storage/cache changes	Caller-supplied revision, warrant unknown; deletion is not disproof	Replaces/removes content and purges source closets	see record	see record	none	none	No automatic evidence-based correction or guaranteed historical rollback established
EPI-RTE-kg-temporal	retention	implemented	EPI-OBJ-kg-triple	truth-apt transformation: acquisition/import	Supplied relation into triples	see record	see record	see record	Persistent insertion	Structural/interval validity, not external fact truth	Admits optional-provenance assertions	see record	see record	none	confidence=1.0 is not acceptance	Underlying predicate semantics remain caller-defined
EPI-RTE-kg-temporal	disposition/acceptance	implemented	EPI-OBJ-kg-triple	truth-apt transformation: acquisition/import	Caller marks ended or supplies successor	see record	see record	see record	Enforcing metadata/state update	Withdrawal/replacement request, not checked criticism	Changes later temporal classification; preserves historical route	see record	see record	none	active_facts is not endorsement	Complete supersede transaction/query boundary not inspected
EPI-RTE-relevance-tags	content transformation	implemented	EPI-OBJ-search-tags	truth-apt transformation: entailed derivation	Distance/timestamp to algorithm-relative tags	see record	see record	see record	Advisory annotation	Formula-relative class/rank only; premises' accuracy unchecked	Hints returned without changing hit order	see record	see record	none	Relevance/freshness is not truth	No calibration evidence
RT-RTE-palace-recall	operational admission/selection/consumption	implemented	EPI-OBJ-drawer-content	truth-apt transformation: non-ampliative reshaping	Retained text to selected context	see record	see record	see record	Ranking and context supply	Retained testimony; no added source warrant	May shape later answers if consumed	see RT-BAP-agent-context	see record	RT-CLM-recall-protocol	Full verbatim delivery not assured	Hydration/truncation prevents completeness inference
EPI-RTE-palace-audit	check/evidence production	implemented	EPI-OBJ-audit-report	truth-apt transformation: entailed derivation	Organization under fixed score formula	see record	see record	see record	Report	Formula result, not proposition truth or measured navigability	Supplies optional threshold consumer	see record	see record	none	none	Analyzer correctness and real utility not measured
EPI-RTE-palace-audit	operational admission/selection/consumption	implemented	EPI-OBJ-audit-report	no content change	Overall score versus requested threshold	see record	see record	see record	Exit status 2 below threshold	No claim accepted or rejected	Signals CLI failure; does not mutate palace	CLI exit, enforcing process result; external caller may gate work	see record	none	none	External CI/maintenance consumption unobserved
EPI-RTE-tunnel-maintenance	disposition/acceptance	implemented	none — navigation proposal	non-truth-apt policy/content update: caller-approved graph navigation changes	Proposed tunnel plan or prune request to stored graph	Caller selects --yes; code rejects stale endpoints and delegates apply/prune	see record	see record	Enforcing graph maintenance	No truth acceptance; inferred associations remain unassessed	Adds/removes navigation links	CLI flag to graph writer, enforcing persistent navigation state	see record	none	Association strength is not warrant	Underlying algorithms and later graph use uninspected
EPI-RTE-retrieval-evaluation	check/evidence production	implemented	EPI-OBJ-retrieval-metrics	truth-apt transformation: entailed derivation	Ranking versus labelled session IDs	see record	see record	see record	Reporting/persistence	Label-relative retrieval success only	Produces report, no demonstrated live-memory acceptance or policy update	see record	see record	EPI-CLM-benchmark-recall	Retrieval is not QA accuracy	Dataset, execution and causal contrast missing

The derivations above are domain-local: an audit formula or benchmark set-membership formula derives a computed property from supplied counts/labels. Neither derives a warranted proposition about the user's world. No assessed row establishes lifecycle integration after evidence-consuming acceptance. This is a bounded finding about these rows, not a searched system-wide absence.

System-claim versus route comparison

Claim ID	Claimed operation or warrant	Claim source ID/anchor and layer	Doctrine/design support	Implemented route IDs	Observed-run support	Causal support and design limits	Supported conclusion	Mismatch/unknown
RT-CLM-recall-protocol	see record	see record	see record	RT-RTE-palace-recall, RT-RTE-hermes-capture, EPI-RTE-checkpoint-disposition	None supplied	Capture drop, near-duplicate rejection and selected/truncated returns are counterexamples to guaranteed complete verbatim recall	Shipped request to search/quote, supported by retained-content delivery mechanisms	Host compliance and answer truth unknown; skill policy not enforceable truth gate
EPI-CLM-benchmark-recall	see record	see record	see record	EPI-RTE-retrieval-evaluation	No executed/reviewed result trace here	Fixed labels define relevance, not truth; reranked/full-set figure and held-out figure have different conditions	Implements bounded retrieval metric; source reports results and explicitly separates QA accuracy	Numbers unreproduced; component attribution and live-loop transfer unsupported
RT-CLM-readonly-boundary	see record	see record	see record	RT-RTE-mcp-dispatch	None supplied	Name-based request gate excludes direct writes and recovery exceptions	Protocol-level operational refusal, not endorsement or universal immutable storage	See record for stronger-claim counterexamples

Bounded conclusion

MemPalace can import caller and host testimony, retain it before any truth acceptance, and deliver selected passages to later agents. Source/session/time metadata preserves some attribution, while the evidence warrant of source propositions remains unknown. Caller-authored assistant text is retained alongside user text; storage does not upgrade either into established knowledge. See RT-RTE-hermes-capture, RT-RTE-mcp-dispatch and RT-RTE-palace-recall.

The material admission criteria differ. Drawer writes have structural, ID and visibility checks. Checkpoints additionally reject near matches at a configurable similarity threshold, with a fail-open path for dedup errors. These criteria target encoding/storage or approximate redundancy, not factual accuracy. A rejected near match can contain a correction, so suppression is consequential without constituting checked semantic equivalence. Explicit replacement/deletion supplies a separate correction capability; cleanup of closets prevents continued selection from stale source summaries, but does not verify the replacement. See EPI-RTE-duplicate-check, EPI-RTE-checkpoint-disposition and EPI-RTE-drawer-revision.

KG assertions differ from ordinary text: code validates interval structure, persists triples and allows caller-requested temporal withdrawal/successor installation. The query wrapper separates active, historical and future results. This gives temporal state operational authority at a later reader, not truth endorsement; optional provenance and default confidence do not establish acceptance. See EPI-RTE-kg-temporal.

Navigation maintenance separately retains a reviewable proposal and uses an operator flag to apply or prune graph links; stale-wing checks constrain operational admission, not the truth of inferred associations. See EPI-RTE-tunnel-maintenance.

Algorithm-relative relevance tags, organization scores and label-relative retrieval metrics are inspectable derivations in narrow formal domains. Their premises are not independently warranted here. Audit has a consequential threshold-to-exit consumer; checkpoint has a check-to-skip consumer; retrieval evaluation produces reports rather than a demonstrated change to live memory or answer policy. These distinct consumers prevent treating all evaluators as one oracle. See EPI-RTE-relevance-tags, EPI-RTE-palace-audit and EPI-RTE-retrieval-evaluation.

Setup classifications can be model/heuristic candidates admitted automatically or edited by people, as in RT-RTE-init-detection. Whether particular outputs preserve inputs, derive facts or conjecture beyond them is not determinable from the supplied route account alone. No assessed route supplies an evidence-consuming decision that accepts an ampliative proposition for a named truth use and scope, then integrates it on that basis. Unassessed extraction, maintenance and evaluator families preclude promoting that local finding into a complete-system absence.

Retained text can change later behavioral context without changing embedding weights. Delivery and policy requests are implemented; exercised reliance, improved capacity, content-directed iterative criticism and exercised self-improvement are independently unestablished in this evidence. Task launch and patch delivery add operational capability and proposal persistence, not product installation or truth warrant. The report assigns no single system-wide epistemic score.

Shared records

Operative objects

EPI-OBJ-drawer-content

Label: Drawer content

Source-native identity: drawer documents text, including physical chunks linked by parent_drawer_id. Representational form: natural language interpreted by agents/humans. Storage substrate: backend collection documents. Distinct from RT-OBJ-task-request and RT-OBJ-patch-artifact: these are conversation/project memory, not coordination events or diffs. Closest supplied route is RT-RTE-mcp-dispatch, whose effects include this object; the route itself is not an object counterpart. SRC-1 mempalace/mcp_server/tools_write.py. Load-bearing storage/chunk evidence follows in EPI-RTE-drawer-revision and the supplied dispatch record.

EPI-OBJ-kg-triple

Label: Knowledge graph triple

Source-native identity: triples relationship with interval, confidence and optional provenance fields. Representational form: symbolic fields with fixed temporal/database consumers; world-relation interpretation depends on caller-defined predicates and readers. Storage substrate: SQLite. Distinct from supplied drawer capture/search routes and coordination objects: temporal relationship rows have separate update and query consumers. SRC-1 mempalace/knowledge_graph.py.

confidence: float = 1.0, source_closet: str = None, source_file: str = None, source_drawer_id: str = None, adapter_name: str = None, --- mempalace/knowledge_graph.py:250-269

EPI-OBJ-search-tags

Label: Search result annotations

Source-native identity: relevance_confidence, recency_rank, is_latest_record. Representational form: symbolic classifications rendered in a tool response for human/model interpretation. Storage substrate: current response dictionary, not durable acquired memory. Distinct from stored drawers; these are transient derived annotations on RT-RTE-palace-recall output. SRC-1 mempalace/mcp_light_server.py; passage in EPI-RTE-relevance-tags.

EPI-OBJ-audit-report

Label: Palace audit report

Source-native identity: audit report's per-layer/overall scores and findings. Representational form: symbolic numeric results plus natural-language diagnostics. Storage substrate: transient dictionary rendered to terminal/JSON; external retention is not established. No supplied object counterpart after comparison with runtime task/patch objects; navigability evaluation is separate. SRC-1 mempalace/palace_audit.py; formula and consumer passages in EPI-RTE-palace-audit.

EPI-OBJ-retrieval-metrics

Label: Retrieval metrics

Source-native identity: session/turn recall_any, recall_all and NDCG results. Representational form: symbolic floats keyed by rank cutoff; storage substrate: in-memory result log and optional JSONL report. No supplied object counterpart; unlike RT-OBJ-task-request, these are benchmark results, not instructions. SRC-1 benchmarks/longmemeval_bench.py; passages in EPI-RTE-retrieval-evaluation.

Routes

EPI-RTE-duplicate-check

Label: Semantic duplicate check Part of: RT-RTE-mcp-dispatch

This is the check operation within MCP dispatch, not its content-retention or disposition part. Endpoints: candidate content → vector query for five hits → metric-specific distance conversion → threshold matches and boolean. Owner: read handler and backend embedder; context/state: current candidate and indexed drawers. Effects: produces a redundancy diagnostic, consumed separately by checkpoint. SRC-1 mempalace/mcp_server/tools_read.py.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: is_duplicate and matches, or vector-disabled/error diagnostic.
• Later read-back: inapplicable — the diagnostic is not persisted as knowledge by this handler; the check reads retained drawers.
• Delegated visibility: inapplicable — this operation launches no worker; external embedding-provider behavior is uninspected.
• Selection predicate: nearest five queried hits whose rounded metric-derived similarity meets caller threshold, default 0.9.
• Invalidation or expiry: current-call result; changes to corpus/model/threshold can change later checks.
• Activation or effect: diagnostic production wired; checkpoint skip consumer inspected separately, actual execution unobserved.
• Evidence limits: top-five and embedding closeness do not establish exhaustive duplicates or semantic equivalence; provider/backend versions differ.

Code proposes the match classification; caller supplies candidate/threshold; software decides it, with no human veto inside the check. Open request mode, no reference-answer oracle. Fixed similarity rules guide the result. On unusable vectors the handler exposes a limitation instead of asserting a trustworthy negative.

similarity = round(_distance_to_similarity(dist, metric), 3) if similarity >= threshold: --- mempalace/mcp_server/tools_read.py

return { "is_duplicate": len(duplicates) > 0, "matches": duplicates, } --- mempalace/mcp_server/tools_read.py

EPI-RTE-checkpoint-disposition

Label: Checkpoint duplicate disposition Part of: RT-RTE-mcp-dispatch

Endpoints: checkpoint item → duplicate diagnostic → skip or tool_add_drawer → aggregate result; optional diary follows. Owner: checkpoint software; context/state: submitted items and shared palace. This differs from EPI-RTE-duplicate-check, which supplies evidence but does not itself reject filing. SRC-1 mempalace/mcp_server/tools_diary.py.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: lists of added, duplicates, errors and optional diary result.
• Later read-back: successfully added items can reach drawer search; skipped item content is not filed by this route.
• Delegated visibility: inapplicable — batch saving launches no isolated agent.
• Selection predicate: valid item fields; positive is_duplicate skips; otherwise invokes drawer add, including on dedup error.
• Invalidation or expiry: checkpoint result is transient; resulting drawers persist until other maintenance.
• Activation or effect: enforcing skip/add branch wired, no supplied checkpoint execution.
• Evidence limits: skip establishes storage rejection only, not falsity or equivalence; each item can fail separately, not one all-or-none batch.

Client proposes content, software decides storage admission, human contribution is unresolved from client identity. No per-item confirmation or answer oracle. Open save mode; configurable threshold and verbatim-retention priority shape disposition. Rejected near matches are not a rollback mechanism; client can choose another write path. No truth acceptance or criticism-based successor selection is established.

dup = tool_check_duplicate(content, threshold=dedup_threshold) if dup.get("is_duplicate"): out["duplicates"].append({"room": room, "matches": dup.get("matches", [])}) continue --- mempalace/mcp_server/tools_diary.py

On a dedup error (genuine index failure — content is guaranteed a

string by the guard above) we still file rather than drop the

memory: verbatim recall is the priority and add_drawer's own

idempotency blocks exact duplicates.

res = tool_add_drawer(wing=wing, room=room, content=content, added_by=resolved_added_by) --- mempalace/mcp_server/tools_diary.py

EPI-RTE-drawer-revision

Label: Drawer replacement and withdrawal Part of: RT-RTE-mcp-dispatch

Endpoints: explicit update/delete request → logical drawer resolution/validation → backend update/upsert/delete → cache invalidation and source-closet purge → result. Owner: write handler; state: old document/chunks/metadata and caller replacement. These are maintenance parts of supplied dispatch, not automatic truth criticism. SRC-1 mempalace/mcp_server/tools_write.py, mempalace/config.py.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: success with drawer/chunk/closet information, not-found/validation error, or backend failure.
• Later read-back: revised content/metadata can reach search/direct reads; removed rows no longer supply those reads if backend mutation succeeds.
• Delegated visibility: inapplicable — maintenance launches no worker.
• Selection predicate: caller-specified drawer ID; logical handle resolves group while a physical chunk delete selects one row; new content/names validated.
• Invalidation or expiry: update sets last_modified; stale chunks removed after replacement upsert; content changes/deletes purge source closets and invalidate overview caches.
• Activation or effect: storage and later-selection changes wired; no observed mutation or answer dependence.
• Evidence limits: backend failure/atomicity not independently checked; WAL intent/result is not a demonstrated undo history or evidence-consuming correctness decision.

Caller proposes replacement/removal, software can veto malformed input/not-found/backend errors. Human authorship/approval is not guaranteed by an MCP client. Open maintenance request, no answer oracle. Fixed validation/chunking/cleanup guidance governs it; explicit later writes can replace content, but automatic truth rollback is not established.

Concrete deterministic trace, deduced without execution: with configured chunk_size=4, input Alice left. passes the shown nonempty/length/null checks and contains no lone surrogate. Add slices it into Alic, e le, ft. with ordered chunk linkage. Concatenating in that order preserves the input proposition; none of the three chunks individually is guaranteed to state it. This supports non-ampliative reshaping, not a new fact. Invalid Unicode sanitization remains a separate fidelity limit.

if not isinstance(value, str) or not value.strip(): raise ValueError("content must be a non-empty string") if len(value) > max_length: raise ValueError(f"content exceeds maximum length of {max_length} characters") if "\x00" in value: raise ValueError("content contains null bytes") return strip_lone_surrogates(value) --- mempalace/config.py

for i in range(0, len(content), chunk_size): chunk_idx = i // chunk_size chunk_ids.append(f"{drawer_id}chunk{chunk_idx:06d}") chunk_docs.append(content[i : i + chunk_size]) --- mempalace/mcp_server/tools_write.py

if content is not None and source_file: closets_deleted = _purge_source_closets(source_file, commit=True) --- mempalace/mcp_server/tools_write.py

col.delete(ids=record["ids"]) _invalidate_overview_caches()

source_file = record["metadata"].get("source_file") closets_deleted = _purge_source_closets(source_file, commit=True) if source_file else 0 --- mempalace/mcp_server/tools_write.py

EPI-RTE-kg-temporal

Label: Knowledge graph temporal maintenance Part of: RT-RTE-mcp-dispatch

Endpoints: caller relation/interval → input and temporal validation → SQLite triple insert or interval end/successor request → later KG query wrapper's active/historical/future grouping. Owner: KG handler and graph code; state: entities, triples and optional provenance. Material part distinct from drawer replacement: stored intervals change temporal query eligibility without deleting historical relationships. SRC-1 mempalace/mcp_server/tools_kg.py, mempalace/knowledge_graph.py.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: triple ID/fact, resolved ended value, superseded value or input/domain error.
• Later read-back: KG query returns stored relationships and temporal categories; timeline is an exposed alternate not traced here.
• Delegated visibility: inapplicable — graph maintenance launches no separate agent.
• Selection predicate: supplied subject/predicate/object; invalidate targets open matching rows; interval checks reject ends preceding starts; supersede supplies old/new object and boundary.
• Invalidation or expiry: invalidate stores end time; supersede requests shared boundary, default current UTC; add does not automatically resolve conflicting alternative objects.
• Activation or effect: add/invalidate database writes and wrapper temporal categorization inspected; supersede handler and boundary preparation wired, complete atomic transaction/as-of guarantee uninspected.
• Evidence limits: callers decide external truth; provenance optional; complete supersede transaction and query SQL not read; no live KG trace.

Caller proposes assertion, ending or successor; software checks structural/interval validity, not evidence for the relationship. No human review required by handlers, no answer oracle; open requests. Fixed temporal schema guides the change. Historical categories preserve some withdrawn state; this is not an established truth-directed criticism loop or automatic rollback. An inverted interval is rejected because it would hide the row from queries, not because its world proposition has been falsified.

if ( valid_from is not None and valid_to is not None and _temporal_end_key(valid_to) < _temporal_start_key(valid_from) ): raise ValueError( f"valid_to={valid_to!r} is before valid_from={valid_from!r}; " "an inverted interval would be invisible to every KG query" ) --- mempalace/knowledge_graph.py

conn.execute( "UPDATE triples SET valid_to=? " "WHERE subject=? AND predicate=? AND object=? AND valid_to IS NULL", (ended, sub_id, pred, obj_id), ) --- mempalace/knowledge_graph.py

"active_facts": active, "historical_facts": historical, "future_facts": future, "facts": results, --- mempalace/mcp_server/tools_kg.py

EPI-RTE-relevance-tags

Label: Light search relevance annotations Part of: RT-RTE-palace-recall

Endpoints: search result → minimum available distance and timestamp order → relevance/recency annotations → light response. Owner: deterministic light wrapper; current response state changes, hit ordering retained. SRC-1 mempalace/mcp_light_server.py.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: enriched response or no-matches status; original search ranking remains.
• Later read-back: inapplicable — wrapper does not persist annotations; response carries previously retained hit text.
• Delegated visibility: inapplicable — no agent is launched by enrichment.
• Selection predicate: existing results; minimum nonmissing distance; descending available filing/creation timestamp for recency.
• Invalidation or expiry: invocation-local; later query or timestamps can yield other annotations.
• Activation or effect: annotations wired into returned envelope; downstream use and improved decisions unobserved.
• Evidence limits: fixed thresholds not calibrated truth or retrieval-success probability; missing/incorrect timestamps defeat stronger freshness conclusions.

Concrete deduction: if the only available distance is 0.34, minimum distance is below 0.35 and the output class is high; 0.60 gives low. This derives a formula-relative tag, not an entailed world proposition. Software computes it; no candidate acceptance or answer oracle occurs. Annotation channel is advisory to response consumers for the current invocation.

distances = [d for d in (_search_distance(item) for item in results_list) if d is not None] if distances: min_dist = min(distances) if min_dist > 0.52: res["relevance_confidence"] = "low" elif min_dist < 0.35: res["relevance_confidence"] = "high" else: res["relevance_confidence"] = "moderate" --- mempalace/mcp_light_server.py

EPI-RTE-palace-audit

Label: Palace navigability audit

No counterpart in the supplied runtime after comparison: unlike dispatch gates and setup detection, this route evaluates stored organization and can change CLI exit status. Endpoints: palace counts/organization analysis → score formula/report → JSON/terminal output → optional fail_under comparator. Owner: audit code and CLI; context/state: current organization snapshot. SRC-1 mempalace/palace_audit.py, mempalace/cli/cmd_audit.py.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: report printed, or exit 2 when nonmissing overall score falls below supplied threshold.
• Later read-back: inapplicable — inspected CLI does not retain report; an external caller may store JSON but that path is uninspected.
• Delegated visibility: inapplicable — audit launches no agent.
• Selection predicate: selected palace; layer availability determines which scores enter mean; threshold only applies when supplied and overall is not None.
• Invalidation or expiry: current snapshot report, no palace mutation by this command.
• Activation or effect: threshold changes process outcome; automated maintenance or downstream gating is unobserved.
• Evidence limits: analyzer internals only selectively inspected; scores encode navigational heuristics, not observed usability or factual truth.

Software evaluates generic-room share, name drift and relationship artifacts/coverage; caller selects failure threshold and can veto subsequent work externally. No expected-answer oracle. Open audit mode; fixed formula is the governing guidance, not a learned revision theory. The CLI returns a signal, not a factual-content rejection or automatic repair.

scores["rooms"] = _clamp(100 * (1 - rooms["generic_share"])) if rooms["total_drawers"] else None --- mempalace/palace_audit.py

present = [s for s in scores.values() if s is not None] scores["overall"] = _clamp(sum(present) / len(present)) if present else None return scores --- mempalace/palace_audit.py

threshold = getattr(args, "fail_under", None) overall = report["scores"]["overall"] if threshold is not None and overall is not None and overall < threshold: sys.exit(2) --- mempalace/cli/cmd_audit.py

EPI-RTE-tunnel-maintenance

Label: Tunnel proposal admission and cleanup

No supplied counterpart after comparison: this is navigation-graph maintenance rather than task patch delivery or drawer capture. Endpoints: current wings/hallways → proposal or prune report → optional user-edited plan → --yes apply/prune branch → saved tunnels. Owner: CLI orchestration and imported tunnel functions; context/state: persistent proposal plan and navigation graph. Only the orchestration was directly inspected. SRC-1 mempalace/cli/cmd_tunnels.py.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: dry-run/proposal report, created/removed count, or stale/invalid-plan error.
• Later read-back: proposal file is reloaded by a later apply invocation; saved tunnels belong to graph navigation, whose downstream internals remain uninspected here.
• Delegated visibility: inapplicable — this CLI route launches no agent.
• Selection predicate: action propose/prune; --yes selects graph modification; apply drops plan rows whose normalized wings no longer exist and refuses if none remain.
• Invalidation or expiry: proposal can be manually edited; stale endpoints filtered on application; prune can remove reported artifacts under the graph writer lock.
• Activation or effect: application and cleanup calls wired; graph-write algorithms, actual changes and later traversal effects unobserved.
• Evidence limits: underlying strength, artifact and application algorithms not inspected; no warrant about real-world associations or measured navigation improvement.

Code proposes a plan; operator can delete unwanted rows and chooses apply. This is human-editable operational admission, not evidence-consuming acceptance of the association's truth. --yes on prune executes the delegated cleanup without a separate dialogue. Open maintenance mode, no answer oracle. Fixed proposal/endpoint rules guide it. No demonstrated rollback or automatic policy-improvement loop follows. Semantic relation of underlying inferred associations is not determinable; the inspected boundary establishes only persistent navigation-state update.

print(f"\n Plan saved to {path}. Delete rows you do not want, then re-run with --yes.") --- mempalace/cli/cmd_tunnels.py

if not fresh: print(f" Every proposed tunnel names a wing that no longer exists ({stale} rows).") print(" Re-run mempalace tunnels propose against the current wings.") sys.exit(1) created = apply_proposal({**plan, "tunnels": fresh}, config=config) --- mempalace/cli/cmd_tunnels.py

with mine_lock(_get_tunnel_file(config)): kept, report = prune_tunnels(_load_tunnels(config), wings) if report["removed"]: _save_tunnels(kept, config) --- mempalace/cli/cmd_tunnels.py

EPI-RTE-retrieval-evaluation

Label: Labelled retrieval evaluation

No counterpart in supplied runtime records: evaluator existence was mentioned, but no evaluator route was declared. Endpoints: benchmark ranking and labelled answer session IDs → session/turn membership metrics → aggregate/per-type reporting and optional JSONL result log. Owner: benchmark harness; context/state: benchmark corpus/rankings/reference IDs. SRC-1 benchmarks/longmemeval_bench.py.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: rank-cutoff metrics and printed summary; optional result file.
• Later read-back: optional persisted report is available to later file readers; no inspected automatic live-policy consumer or demonstrated later reliance.
• Delegated visibility: inapplicable — inspected metric calculation delegates no agent; other retrieval/reranking modes have uninspected provider behavior.
• Selection predicate: top-k ranking IDs versus reference IDs; session-level corpus IDs collapse to session identities; turn reference set includes all corpus IDs from answer sessions.
• Invalidation or expiry: results belong to dataset/mode/configuration; no automatic expiration or product rollback in the inspected save path.
• Activation or effect: report production/persistence wired; route not executed, no observed product change.
• Evidence limits: expected-session membership is not end-to-end answer correctness; turn labels inherit session attribution, not independently annotated minimal evidence; dataset and retained result files not inspected.

Harness supplies the ranking; dataset supplies the reference-outcome oracle (answer_sids); code computes metrics. This is bounded benchmark evaluation, not open live-answer acceptance. Human reference construction is uninspected. Report saving does not itself select a production successor, reject memories or establish an evidence-responsive improvement loop. Concrete deduction: rankings [0], corpus IDs [s1], correct IDs {s1}, k=1 gives both recalls 1; replacing correct IDs with {s2} gives both 0. The checked domain is labelled retrieval membership only.

top_k_ids = set(corpus_ids[idx] for idx in rankings[:k]) recall_any = float(any(cid in top_k_ids for cid in correct_ids)) recall_all = float(all(cid in top_k_ids for cid in correct_ids)) ndcg_score = ndcg(rankings, correct_ids, corpus_ids, k) return recall_any, recall_all, ndcg_score --- benchmarks/longmemeval_bench.py

session_level_ids = [session_id_from_corpus_id(cid) for cid in corpus_ids] session_correct = answer_sids --- benchmarks/longmemeval_bench.py

if out_file: with open(out_file, "w") as f: for entry in results_log: f.write(json.dumps(entry) + "\n") print(f" Results saved to: {out_file}") --- benchmarks/longmemeval_bench.py

Claims

EPI-CLM-benchmark-recall

Label: Reported benchmark recall

Claimed operation: README reports raw LongMemEval R@5 96.6%, hybrid held-out 98.4%, and rerank full-set ≥99%, while warning against equating retrieval recall with QA accuracy and against test-specific tuning. Source: SRC-1 README.md, doctrine/design and attributed reported results, not an inspected run trace. Distinct from RT-CLM-recall-protocol: numeric evaluator claims are not the agent's recall policy. The registered source layer and this inspection support attribution and evaluator implementation, not independent result authentication. EPI-RTE-retrieval-evaluation supplies the bounded metric interpretation.

• doctrine conclusion status: claimed
• operation conclusion status: uninspected

| Raw (semantic search, no heuristics, no LLM) | 96.6% | None | | Hybrid v4, held-out 450q (tuned on 50 dev, not seen during training) | 98.4% | None | | Hybrid v4 + LLM rerank (full 500) | ≥99% | Any capable model | --- README.md

metrics on different splits, and placing retrieval recall next to end-to-end QA accuracy is not an honest comparison. See each project's own research page for their published numbers. --- README.md