MemoryOS epistemic report
Type: agentic-system-analyses/types/agentic-system-epistemic-report.md
Source-and-claim boundary
The question is whether MemoryOS's memory-assisted dialogue and maintenance routes produce warranted content, rather than merely retaining, interpreting and reusing it. Scope and exclusions are see boundary; source identity and evidence layers are see Source register. All source references below are SRC-1 at that pin. No target execution was performed. Implementation supports possible routes, not their activation or benefit.
The consequential system claim is see RT-CLM-dialogue-benefit. Prompt demands for factual extraction and evidence-grounded preference inference are assessed separately below; a request for factuality is not a completed factual check.
| Entry point or operation | Source path | Coverage and record |
|---|---|---|
| Python add/get_response, summary, continuity, metadata and profile maintenance | memoryos-pypi/memoryos.py, memoryos-pypi/updater.py, memoryos-pypi/utils.py, memoryos-pypi/prompts.py |
Direct reads of maintenance admission, updater, utility transformations and prompts; remaining orchestration/retrieval/storage supplied by RT-RTE-dialogue-response, RT-RTE-memory-admission, RT-RTE-profile-analysis. |
| MCP add/retrieve/profile, stdio dispatch | memoryos-mcp/server_new.py |
Decorator/definition search checked against supplied RT-RTE-mcp-access; bundled helper details rely on that record, not presumed universal equivalence. |
| Playground chat/import/force/clear/state | memoryos-playground/memdemo/app.py |
Direct route search; supplied RT-RTE-playground-control. |
| Playground personality-analysis display | memoryos-playground/memdemo/app.py |
Direct read; EPI-RTE-personality-display adds a material interpretive operation not separately traced in runtime. |
| ChromaDB keyword/profile transformations and persistent caller | memoryos-chromadb/utils.py, memoryos-chromadb/mid_term.py |
Direct utility read and method search; other insertion, retrieval and persistence findings supplied by RT-RTE-chroma-dialogue. Deep insertion/merge internals remain unassessed; prevents a variant-complete negative about lineage or acceptance. |
| Evaluation import, generation and offline scoring | eval/main_loco_parse.py, eval/dynamic_update.py, eval/evalution_loco.py |
Direct scorer read and updater search; other workflow findings supplied by RT-RTE-locomo-evaluation. Evaluation storage internals and full prompt bodies remain unassessed; no system-complete absence claim. |
| Docker, demos, model components and remaining documentation | Dockerfile, README.md, memoryos-pypi/test.py |
Supplied runtime account and RT-CMP-openai-client, RT-CMP-sentence-transformer, RT-CMP-bgem3-flag-model; no additional CLI dispatcher or autonomous hook inferred. |
Missing execution evidence prevents concluding that a particular extraction, admission, consumer influence or improvement occurred. Missing provider internals, resolved model versions and weights prevent identifying parametric propositions or their warrant; truth-apt objects inside those weights are not determinable. Available but uninspected variant code is an inspection limit, not an access gap. The excluded paper and unfrozen deployments support no extra claim here. No supplied fact requires supersession; the display operation supplements a valid runtime container.
Epistemic-object inventory
| Object ID | Candidate truth-apt content or none | Claimed role | Evidence source ID and local anchor | Gap/limit |
|---|---|---|---|---|
| EPI-OBJ-dialogue | User assertions and assistant answers, when proposition-bearing | see record | SRC-1 memoryos-pypi/updater.py |
Imperatives and social turns need not be truth-apt; assertions lack independent source certification. |
| EPI-OBJ-summaries | Topic summaries and continuity metadata | see record | SRC-1 memoryos-pypi/utils.py |
Requested compression does not prove semantic preservation. |
| EPI-OBJ-user-profile | Preference/trait levels, contextual explanations | see record | SRC-1 memoryos-pypi/prompts.py |
A possible inference about a person, not an independently measured trait. |
| EPI-OBJ-user-knowledge | Extracted personal facts | see record | SRC-1 memoryos-pypi/utils.py |
No item-level evidence-consuming truth acceptance shown. |
| EPI-OBJ-assistant-knowledge | Claims about assistant actions/capabilities | see record | SRC-1 memoryos-pypi/prompts.py |
A statement of capability in dialogue does not demonstrate that capability. |
| RT-OBJ-generation-prompts | Instructions primarily; supplied content may assert facts | see record | SRC-1 memoryos-pypi/prompts.py |
Prompt instruction and its factual payload have different roles. |
| EPI-OBJ-personality-traits | Displayed dimension/level assignments | see record | SRC-1 memoryos-playground/memdemo/app.py |
Parser can manufacture a positive-looking assignment from a negation. |
| RT-CMP-openai-client, RT-CMP-sentence-transformer, RT-CMP-bgem3-flag-model | Not determinable within opaque distributed-parametric state; vectors used for relevance are not truth verdicts | see records | SRC-1 memoryos-pypi/utils.py |
Encoding calls are inspected/supplied, weights and provider processing are not. |
Authority-route ledger
Fields owned by cited records are not restated: see delegates those fields to the declaration. This ledger supplies the epistemic classification. Rows sharing an ID separate its consequential functions, not new referents. All implemented statuses mean implementation evidence only; no observed-run result is claimed.
| 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-memory-admission | content transformation | implemented | EPI-OBJ-dialogue | truth-apt transformation: acquisition/import | Supplied dialogue enters memory | see record | see record | see record | see record | Source warrant unknown; retains assertions, not established facts | see record | see record | SRC-1 memoryos-pypi/memoryos.py |
none | none | External/imported assistant text need not originate from this response model. |
| RT-RTE-memory-admission | content transformation | implemented | EPI-OBJ-summaries | truth-apt transformation: indeterminate | Dialogue to summaries/meta-summary | see record; generation, not an independent verifier | see record | see record | see record | Neither entailment nor semantic preservation established | see record | see record | SRC-1 memoryos-pypi/updater.py, memoryos-pypi/utils.py |
none | none | Reshaping versus conjectural content needs actual input/output and semantic checking. |
| RT-RTE-memory-admission | check/evidence production | implemented | EPI-OBJ-summaries | no content change | Topic continuity of adjacent pages; JSON syntax of summary | Model true/false judgment over supplied dialogue; JSON parser over syntax | Overflow processing; see record | Boolean continuity; parsed array or empty fallback | Branch selection only | Supports an attributed continuity judgment or parse success, not source truth | Updater can link pages and propagate metadata | see record | SRC-1 memoryos-pypi/utils.py, memoryos-pypi/updater.py |
none | none | Exact true test means true. is false; no independent continuity oracle. |
| RT-RTE-memory-admission | retention | implemented | EPI-OBJ-dialogue, EPI-OBJ-summaries | no content change | Admission to short/mid-term state | see record | see record | see record | see record | Retention grants availability, not evidence-consuming acceptance | see record | see record | SRC-1 memoryos-pypi/updater.py |
none | none | No lifecycle integration inferred from grouping or persistence. |
| RT-RTE-profile-analysis | content transformation | implemented | EPI-OBJ-user-profile | truth-apt transformation: indeterminate | Dialogue plus old profile to revised trait claims | see record; prompt asks inference from evidence, not a separate test | see record | see record | see record | Candidate interpretations only | see record | see record | SRC-1 memoryos-pypi/prompts.py, memoryos-pypi/utils.py |
none | none | Trait inference can be ampliative, but no actual non-entailed output established; classification stays indeterminate. |
| RT-RTE-profile-analysis | content transformation | implemented | EPI-OBJ-user-knowledge, EPI-OBJ-assistant-knowledge | truth-apt transformation: indeterminate | Dialogue to concise extracted lines | see record | see record | see record | see record | Requested factual extraction, not certified preservation | see record | see record | SRC-1 memoryos-pypi/utils.py, memoryos-pypi/prompts.py |
none | none | Separate user and assistant consumers remain below. |
| RT-RTE-profile-analysis | check/evidence production | implemented | EPI-OBJ-user-profile, EPI-OBJ-user-knowledge, EPI-OBJ-assistant-knowledge | no content change | Output nonemptiness, sentinel strings, profile length | Deterministic string checks; domain is output shape, not truth | Before writes; see record | Pass or skip | Guards control write eligibility | Only shape/length license; no factual criterion | see record | see record | SRC-1 memoryos-pypi/memoryos.py |
none | none | Error sentinel is longer than thirty characters and not none: it can pass the profile guard. |
| RT-RTE-profile-analysis | disposition/acceptance | implemented | EPI-OBJ-user-profile, EPI-OBJ-user-knowledge, EPI-OBJ-assistant-knowledge | no content change | Guarded content becomes eligible retained context | see record | see record | see record | see record | Operational admission only; not acceptance of truth under a named evidential criterion | see record | see record | SRC-1 memoryos-pypi/memoryos.py |
none | none | Marking pages analyzed records processing, not warranted endorsement. |
| RT-RTE-profile-analysis | retention | implemented | EPI-OBJ-user-profile, EPI-OBJ-user-knowledge, EPI-OBJ-assistant-knowledge | no content change | Profile replacement and fact append | see record | see record | see record | see record | Retained candidate ledger, not established lifecycle integration | see record | see record | SRC-1 memoryos-pypi/memoryos.py |
none | none | Old profile feeds later inference, enabling persistence of unsupported claims too. |
| RT-RTE-dialogue-response | operational admission/selection/consumption | implemented | EPI-OBJ-dialogue, EPI-OBJ-user-profile, EPI-OBJ-user-knowledge | no content change | Context selection and delivery | Similarity thresholds/ranking; full profile and recent history bypass ranking; see record | see record | see record | see record | Relevance selection grants no truth license | see record | see RT-BAP-user-context | SRC-1 memoryos-pypi/retriever.py, memoryos-pypi/memoryos.py |
RT-CLM-dialogue-benefit | none | Access/heat is salience, not endorsement or temporal truth. |
| RT-RTE-dialogue-response | operational admission/selection/consumption | implemented | EPI-OBJ-assistant-knowledge | no content change | Assistant facts selected into system message | see record | see record | see record | see record | No independent warrant from higher-priority placement | see record | see RT-BAP-assistant-context | SRC-1 memoryos-pypi/memoryos.py |
RT-CLM-dialogue-benefit | none | Operationally stronger channel need not imply epistemically stronger provenance. |
| RT-RTE-dialogue-response | content transformation | implemented | EPI-OBJ-dialogue | truth-apt transformation: indeterminate | Query/context to answer | see record | see record | see record | see record | No external answer oracle in open dialogue | see record | see record | SRC-1 memoryos-pypi/memoryos.py |
RT-CLM-dialogue-benefit | none | Model may reshape, derive, conjecture or produce non-truth-apt text; no output supplied to distinguish. |
| EPI-RTE-output-cleaning | content transformation | implemented | EPI-OBJ-dialogue | truth-apt transformation: non-ampliative reshaping | Tagged response text | see record | see record | see record | see record | Deletion preserves no warranty of whole-response meaning | see record | see record | SRC-1 memoryos-pypi/utils.py |
none | none | Explicit witness deletes a proposition; no evidence-consuming acceptance occurs. |
| RT-RTE-mcp-access | operational admission/selection/consumption | implemented | EPI-OBJ-dialogue, EPI-OBJ-user-profile, EPI-OBJ-user-knowledge, EPI-OBJ-assistant-knowledge | no content change | Tool result data to host | see record | see record | see record | see record | Host receives candidate context, not adjudicated knowledge | see record | see record | SRC-1 memoryos-mcp/server_new.py |
none | none | Host's subsequent reliance/channel outside boundary. |
| RT-RTE-chroma-dialogue | operational admission/selection/consumption | implemented | EPI-OBJ-summaries, EPI-OBJ-user-profile | no content change | Keyword ranking and profile consumption | see record; JSON parsing or raw-text wrapper does not check truth | see record | see record | see record | Availability and relevance only | see record | see record | SRC-1 memoryos-chromadb/utils.py, memoryos-chromadb/retriever.py |
none | none | Same warrant limits do not establish identical variant admission or persistence. |
| RT-RTE-playground-control | lineage/freshness/recovery | implemented | EPI-OBJ-dialogue, EPI-OBJ-user-profile | no content change | Import, forced analysis and explicit clear | see record | see record | see record | see record | Clear is deletion, not claim-specific retraction or revised warrant | see record | see record | SRC-1 memoryos-playground/memdemo/app.py |
none | none | Eviction/reset controls future availability, not falsehood classification. |
| EPI-RTE-personality-display | content transformation | implemented | EPI-OBJ-personality-traits | truth-apt transformation: ampliative conjecture | Profile prose to displayed level assignment | see record | see record | see record | see record | Concrete negation counterexample defeats entailment | see record | see record | SRC-1 memoryos-playground/memdemo/app.py |
none | none | Symbolic implementation does not make this a valid deduction. |
| RT-RTE-locomo-evaluation | check/evidence production | implemented | EPI-OBJ-dialogue | no content change | Benchmark answer lexical overlap with supplied reference | see record; token-set F1 domain only | see record | see record | see record | Licenses lexical overlap, not factual equivalence or causal component effect | see record | see record | SRC-1 eval/evalution_loco.py |
RT-CLM-dialogue-benefit | none | Offline printed aggregates have no inspected admission, rollback or policy-update consumer. |
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 |
|---|---|---|---|---|---|---|---|---|
| RT-CLM-dialogue-benefit | see record | SRC-1 README.md; doctrine/design, reported operation |
see record | RT-RTE-dialogue-response, RT-RTE-profile-analysis, RT-RTE-locomo-evaluation | see record | see record | Memory-conditioned response and offline F1 machinery are wired | Personalized coherence, numerical gains, BLEU improvement and component causation unverified. |
The extraction prompts request concise factual statements and preference inference from conversation (SRC-1 memoryos-pypi/prompts.py; see EPI-OBJ-user-profile, EPI-OBJ-assistant-knowledge). These are proposal guidance, not an additional public claim that every output has passed a truth test. The guard's permissive error-sentinel case is a concrete counterexample to treating successful admission as successful analysis.
Bounded conclusion
MemoryOS imports dialogue with unknown source warrant, retains original pages as well as model interpretations, and supplies selected material to subsequent models or an MCP host. Model-generated summaries, extracted facts, profiles and answers have indeterminate semantic relations without actual candidates and checking: neither their labels nor prompts establish entailed derivation, novelty or preservation. The deterministic output cleaner proves that intervening deletion occurs; the Playground parser supplies a concrete ampliative edge even without model inference.
Admission criteria in the inspected JSON maintenance route are nonempty/non-sentinel content and a thirty-character profile minimum. Their intended use is persistent dialogue context, not truth certification. Profile replacement and knowledge append are software decisions over model proposals, not per-item human approvals. Their consumer channels and horizons are see RT-BAP-user-context, RT-BAP-assistant-context. User and assistant facts have different downstream authority channels; neither acquires warrant merely by being used. ChromaDB's parsed-or-wrapped profile route has a distinct, weaker shape requirement, not the JSON variant's guard.
Continuity judgments, summary JSON parsing, similarity/heat selection, admission filters and benchmark scoring are real checks, each with a limited target. None of these inspected consumers turns the offline F1 score into content acceptance, source-truth verification or method revision. As a concrete implementation deduction from SRC-1 eval/evalution_loco.py, Alice left and Alice Alice left yield identical token sets and F1 1; Alice left and Alice never left yield precision 1, recall 2/3 and F1 0.8 despite opposed assertions. Thus the implemented lexical evaluator is not a truth oracle. No claim is made about all possible uninspected variant checks.
No evidence-consuming acceptance of a produced ampliative proposition, nor subsequent lifecycle integration after such acceptance, is established by these inspected maintenance paths. Operational organization of candidates is established separately. Deletion, LFU/capacity expiry, raw-text fallback and analyzed markers alter availability or processing; they do not constitute content-directed criticism with blame or evidence of successful correction. The offline evaluator is a counterexample to an unqualified claim that MemoryOS never evaluates anything, while the absence of its consequential update consumer blocks treating it as runtime warrant or an exercised improvement loop.
Trace-fed profile/fact updates can shape future dialogue through retained natural-language state. They do not train the inspected model components. Improved future capacity, exercised self-improvement, reliable model compliance and component benefits remain unsupported without execution and attribution evidence. Preference-model guidance is explicit, and retained profiles feed later updates, but no supplied content-directed criticism establishes a theory-builder cycle. These limits do not negate the independent implemented acquisition, retention, selection and behavioral-authority paths.
Shared records
Operative objects
EPI-OBJ-dialogue
Label: Dialogue assertions and answers
Source-native identity: QA user_input and agent_response, later page fields. Natural-language turns inside symbolic dictionaries; short/mid-term JSON substrate in the principal variant. Evidence: SRC-1 memoryos-pypi/updater.py; supplied RT-RTE-memory-admission gives admission, persistence and read-back. Distinct from RT-OBJ-generation-prompts: accumulated payload, not shipped instructions. No supplied operative-object counterpart after comparison.
"user_input": qa_pair.get("user_input", ""), "agent_response": qa_pair.get("agent_response", ""), ---
memoryos-pypi/updater.py
EPI-OBJ-summaries
Label: Topic and continuity summaries
Source-native identity: theme_summary, page meta_info. Natural-language summaries and symbolic keyword lists in mid-term storage. Role: session grouping/selection and continuity context. Evidence: SRC-1 memoryos-pypi/updater.py, memoryos-pypi/utils.py. Distinct from the raw dialogue object and static RT-OBJ-generation-prompts; closest supplied mechanisms are RT-RTE-memory-admission, not duplicate object declarations.
theme_summary = summary_item.get("content", "General summary of recent interactions.") theme_keywords = summary_item.get("keywords", []) ---
memoryos-pypi/updater.py
EPI-OBJ-user-profile
Label: User preference profile
Source-native identity: updated_user_profile; natural-language trait assessments stored as user long-term profile text in JSON. Role: background conditioning and input to later profile revision. Evidence: SRC-1 memoryos-pypi/prompts.py, memoryos-pypi/memoryos.py; closest supplied record RT-RTE-profile-analysis is its producer container, not an object counterpart. Distinct from RT-OBJ-generation-prompts, the static consumption instructions.
- If reflected, determine the user's preference level: High / Medium / Low, and briefly explain the reasoning, including time, people, and context if possible.
- If the dimension is not reflected, do not extract or list it. ---
memoryos-pypi/prompts.py
EPI-OBJ-user-knowledge
Label: User private knowledge lines
Source-native identity: extracted private text admitted by add_user_knowledge. Natural-language facts held in symbolic entries with embeddings; long-term JSON in the principal variant. Role: query-selected user background. Evidence SRC-1 memoryos-pypi/utils.py, memoryos-pypi/memoryos.py; supplied producer RT-RTE-profile-analysis and consumer RT-BAP-user-context. No supplied operative-object counterpart after comparison; different consumer from assistant facts.
if new_user_private_knowledge and new_user_private_knowledge.lower() != "none": for line in new_user_private_knowledge.split('\n'): if line.strip() and line.strip().lower() not in ["none", "- none", "- none."]: self.user_long_term_memory.add_user_knowledge(line.strip()) ---
memoryos-pypi/memoryos.py
EPI-OBJ-assistant-knowledge
Label: Assistant action and capability lines
Source-native identity: extracted assistant_knowledge admitted by add_assistant_knowledge. Natural-language claims in embedded symbolic entries, assistant-specific long-term JSON in the principal variant. Role: later system-message persona/knowledge. Evidence SRC-1 memoryos-pypi/prompts.py; supplied RT-RTE-profile-analysis, RT-BAP-assistant-context. No supplied operative-object counterpart; source and consumer distinguish it from user facts and static prompts.
Assistant knowledge: explicit statements about what the assistant did, provided, or demonstrated ---
memoryos-pypi/prompts.py
EPI-OBJ-personality-traits
Label: Parsed personality display assignments
Source-native identity: extracted_traits, returned as personality_analysis. Symbolic category/dimension/level dictionaries carrying proposition-like trait assignments; transient HTTP JSON substrate, not a retained memory write. Evidence SRC-1 memoryos-playground/memdemo/app.py. Closest supplied RT-RTE-playground-control describes the application, not this operative object. Distinct from profile prose because the parser omits qualifications and chooses levels by lexical occurrence.
extracted_traits[category].append({ 'dimension': dimension, 'level': level }) ---
memoryos-playground/memdemo/app.py:325-328
Routes
EPI-RTE-output-cleaning
Label: Completion output cleaning
Part of: RT-RTE-dialogue-response
This part is the response client's deterministic text-deletion stage; the parent remains valid. Endpoints: raw completion → clean_reasoning_model_output → cleaned return → parent admission. Owner: OpenAIClient. SRC-1 memoryos-pypi/utils.py supplies implementation evidence. The same utility also wraps maintenance calls; no claim of identical cleaning in all evaluation helpers follows. The transformation removes tagged spans rather than evaluating their assertions. For concrete input <think>Alice left.</think>Bob stayed., regex deletion yields Bob stayed.; whitespace cleanup and strip leave it unchanged. This is non-ampliative deletion, not acceptance, evidence production or guaranteed meaning preservation. No human decision, reference answer or quality oracle participates.
- implementation conclusion status: wired
- operation conclusion status: uninspected
- Immediate return: cleaned completion text.
- Later read-back: parent's admitted cleaned answer can enter later memory contexts; removed span is not admitted by this route.
- Delegated visibility: provider produced raw text; caller and subsequent memory admission receive the cleaned result.
- Selection predicate: regex matches literal lowercase think-tag spans, non-greedily across newlines; excess blank lines and boundary whitespace are removed.
- Invalidation or expiry: inapplicable — no independently retained cleaner state; parent memory expiry controls admitted text.
- Activation or effect: deletion is implementation-deduced for the concrete input; no observed model response supplied.
- Evidence limits: no source-truth check, tag correctness check or semantic-preservation proof; unclosed tags do not match this pattern.
cleaned_text = re.sub(r'
.*? ', '', text, flags=re.DOTALL) ---memoryos-pypi/utils.pycleaned_content = clean_reasoning_model_output(raw_content) return cleaned_content ---
memoryos-pypi/utils.py
EPI-RTE-personality-display
Label: Playground personality display parsing
Part of: RT-RTE-playground-control
This part is the application's profile-to-display interpretation, not a replacement for the supplied admin-control container. Endpoints: POST /personality_analysis → retained raw profile → parse_personality_traits → JSON assignments. Owner: Flask callback and deterministic parser. Evidence SRC-1 memoryos-playground/memdemo/app.py, implementation. The model/caller supplies profile prose upstream; code chooses display labels; user requests display but does not approve each label. No truth oracle or improvement trigger. The only callback vetoes are missing session/profile and exceptions, not trait accuracy. No retained update, acceptance or rollback occurs.
Concrete source deduction: input Extraversion is not High. yields no regex pair. The fallback scans High before Medium and Low, finds the dimension substring, and emits Extraversion / High under Psychological Model. The input denies the displayed proposition, so this transformation is demonstrably not entailed reshaping. The classification is ampliative conjecture in the contract's semantic sense, not a claim that this parser intentionally reasons. A positive input such as Extraversion (High) can also produce a faithful assignment; the counterexample bounds reliability rather than asserting every assignment is wrong.
- implementation conclusion status: wired
- operation conclusion status: uninspected
- Immediate return: HTTP JSON
personality_analysiscategory/dimension/level lists. - Later read-back: inapplicable — parser result is returned, not written into memory by this endpoint.
- Delegated visibility: browser/caller receives assignments; no provider call from this callback.
- Selection predicate: session exists and profile is not empty/none-like; regex pair matching plus case-insensitive level/dimension substring fallback, suppressing exact-dimension duplicates.
- Invalidation or expiry: inapplicable — result is transient and recomputed on request; source profile can be replaced by other routes.
- Activation or effect: implemented presentation of possibly amplified trait claims; no supplied browser trace demonstrates display or reliance.
- Evidence limits: detects textual patterns, not negation or warranted preference levels; no evidence of subsequent model consumption of the returned assignments.
personality_analysis = parse_personality_traits(user_profile) ---
memoryos-playground/memdemo/app.pyfor level in ['High', 'Medium', 'Low']: if level.lower() in line.lower(): # Try to extract the dimension name for category, dimensions in categories.items(): for dimension in dimensions: if dimension.lower() in line.lower(): ---
memoryos-playground/memdemo/app.py