Context Constitution epistemic report

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

Source-and-claim boundary

The question is how this published context-management policy proposes to turn experience into content relied on later, and what checks license that reliance. Scope and exclusions: see boundary; evidence identity and layers: see its Source register. This is a complete document artifact, not an inspected Letta runtime. All direct reads below are SRC-1 at the reviewed commit. No source, execution or causal comparison is added.

Direct inspection covered the full README, Constitution, Affordances and whitespace workflow. The supplied runtime account was checked against these documents rather than treated as an inspection limit. No shipped CLI dispatch, registered runtime tools, executable prompt loader, agent hooks or evaluator implementation exists in the tracked tree described by the boundary. Affordances describes operations, not their schemas or dispatch. The publication workflow is executable but outside the selected agent mechanism. Its evaluator tests Markdown trailing spaces, not knowledge, memory quality or improved action. There is therefore no deterministic agent-content transformation to trace through inspected code.

Entry point or operation	Source path	Coverage and records	Conclusion limit
Public entry and prompting/training use	README.md	Direct read; RT-OBJ-context-constitution, RT-CLM-token-learning	Internal use is reported, not an observed deployed prompt or training process.
Durable correction, generalization and self-assessment	constitution/CONSTITUTION.md	Direct read; EPI-OBJ-durable-learnings, EPI-RTE-update-assessment, EPI-CLM-general-learning	No particular candidate, evidence-consuming acceptance or improvement episode supplied.
MemFS, Git propagation, direct memory tools and human edits	constitution/AFFORDANCES.md	Direct read; RT-RTE-memory-admission	Push is an operational propagation condition, not a semantic check; alternate tools have unspecified admission.
External/environment context, indexes, skills and history search/list	constitution/CONSTITUTION.md; constitution/AFFORDANCES.md	Direct reads; RT-RTE-context-disclosure	Selection and delivery do not endorse source content; scripts and arbitrary programmatic operations have no shipped executor to inspect.
Automatic history logging and compaction	constitution/CONSTITUTION.md; constitution/AFFORDANCES.md	Direct reads; RT-RTE-history-compaction, EPI-OBJ-conversation-summary	No stored trace, summarizer, lineage check or summary acceptance result.
Pruning, externalization, correction and withdrawal	constitution/CONSTITUTION.md	Direct read; EPI-RTE-pruning-assessment, RT-RTE-memory-admission	Conservative policy and retrieval confidence do not establish semantic rejection or reliable recovery.
Recall, Reflection, Defragmentation and self-conversations	constitution/AFFORDANCES.md	Direct read; RT-RTE-subagent-context	Roles differ; no return schema, reviewer decision, merge rule or later reliance trace.
Underlying model and model replacement	constitution/CONSTITUTION.md; README.md	Direct reads; RT-CMP-underlying-model	No exact model, accessible weights, parameter-update evidence or cross-model continuity test.
Publication check	.github/workflows/check-whitespace.yml	Direct read; excluded from agent knowledge evaluation as in boundary	Cannot support semantic correctness, policy compliance or learning.

Missing harness implementations prevent implementation findings about synchronization, compaction, permission checks and delivered prompts. Missing candidate memories and input/output traces prevent individuating an actual truth-apt learning, deciding its entailment relation, or establishing acceptance and activation. Missing comparisons prevent benefit, transfer and component-effect attribution. These limits do not erase the positive policy routes. Public learning claims are RT-CLM-token-learning and the narrower transfer requirement EPI-CLM-general-learning. No relevant declared operation is left unassessed at this document boundary; no claim of whole-runtime coverage follows.

Epistemic-object inventory

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source ID and local anchor	Gap/limit
RT-OBJ-context-constitution	Explicit explanatory claims about context, identity and learning; imperatives themselves are policy, not established facts	see record	SRC-1, constitution/CONSTITUTION.md	Published claims are localized; their delivery and meaning-dependent runtime use are not supplied.
EPI-OBJ-durable-learnings	Corrected assumptions, user preferences and environment knowledge can be truth-apt over the user/environment and time concerned; behavioral guidance and persona choices need not be	see record	SRC-1, constitution/CONSTITUTION.md	No actual retained instance: individual propositions, scope and source warrant not determinable.
EPI-OBJ-conversation-summary	Claims about prior events and extracted facts can be true or false about the source conversations	see record	SRC-1, constitution/AFFORDANCES.md	No before/after content, summarizer or fidelity test; summary label alone does not establish semantic preservation.
RT-RTE-context-disclosure	Retrieved messages, files and skills can carry factual assertions; indexes carry discovery cues; exact content not determinable	see record for its operative inputs	SRC-1, constitution/AFFORDANCES.md	Source ownership and existence are not source truth. Environment context and agent-owned memory have different provenance.
RT-CMP-underlying-model	No localized truth-apt proposition individuated in the distributed weights from accessible evidence	see record	SRC-1, constitution/CONSTITUTION.md	Parameters, update behavior and mutable endpoint unresolved; no inference about internal criticism or stored truths.

Authority-route ledger

Each route citation supplies its generic identity, progression and record-owned fields. The following fields assess separate consequential functions; the architectural status is about the agent route, not about publication of its instructions.

Route ID: RT-RTE-context-disclosure Route function: content transformation Architectural status: doctrine only Object/candidate ID: see cited record's external files, skills and history inputs. Content/update relation: truth-apt transformation: acquisition/import; source warrant unknown, not granted by retrieval. Transition or check target: External context enters the current inference window. Evaluator/condition and domain: Agent relevance judgment over the current task, not a truth evaluator. Activation and timing: On task demand after cue and metadata inspection; operation unobserved. Possible or observed result: Requested fuller content available; exact result see record. Implemented force: None established in this repository; declared model-visible supply. Epistemic authority and scope: Access to source content only; no independent endorsement. Operational authority: behavior permitted, blocked, or changed: Selected skills or records can guide the current task. Behavioral-authority path: consumer, channel, force, horizon: Foreground model; retrieved text/tool return; advisory natural-language guidance; current task and any later retrieval. Evidence source ID and local anchor: SRC-1, constitution/CONSTITUTION.md, constitution/AFFORDANCES.md; passages see record. Claim IDs or none: RT-CLM-token-learning. Mismatch marker or none: Retrieval is not acceptance. Gap/limit: No source-verification contract, exact payload or behavioral dependence supplied.

Route ID: RT-RTE-memory-admission Route function: content transformation Architectural status: doctrine only Object/candidate ID: EPI-OBJ-durable-learnings. Content/update relation: truth-apt transformation: indeterminate; the instruction covers copying corrections, reshaping experience, and generalizing it, but no concrete input/output pair distinguishes acquisition, non-ampliative reshaping, entailed derivation or ampliative conjecture. Persona/guideline edits separately permit non-truth-apt policy updates. Transition or check target: Experience/current context to proposed durable memory and prompts. Evaluator/condition and domain: Agent decides durability and general usefulness; not an inspected semantic transformation. Activation and timing: Incremental, deliberate changes after experience; see record. Possible or observed result: Revised assumptions, preferences, identity or guidance; no observed instance. Implemented force: None established; proposed text change. Epistemic authority and scope: Candidate status only; no warranted premises or checked interpretation supplied. Operational authority: behavior permitted, blocked, or changed: Proposed changes can enter the memory admission route. Behavioral-authority path: consumer, channel, force, horizon: Editing agent, then later model; memory/prompt text; proposed behavioral guidance; across invocations if propagated and consumed. Evidence source ID and local anchor: SRC-1, constitution/CONSTITUTION.md; RT-RTE-memory-admission supplies the verification qualification. Claim IDs or none: RT-CLM-token-learning, EPI-CLM-general-learning. Mismatch marker or none: Generalization objective does not itself establish an accepted ampliative claim. Gap/limit: Neither non-entailment nor entailment can be shown without actual candidate content; fluency or persistence would not settle either.

Route ID: EPI-RTE-update-assessment Route function: check/evidence production Architectural status: doctrine only Object/candidate ID: EPI-OBJ-durable-learnings. Content/update relation: no content change. Transition or check target: Whether memory/prompts reduce repeated mistakes and improve future help to this user. Evaluator/condition and domain: Agent self-questions and observation over time; domain is usefulness, remembered preferences and corrective effects, not a universal truth test. Activation and timing: Memory-quality reflection and later observation after updates. Possible or observed result: A judgment that prompts need correction, refinement or continued use; no recorded result. Implemented force: None established; evaluative policy. Epistemic authority and scope: Could supply task/user-specific feedback if performed; no disclosed criterion, measured contrast or acceptance decision. Operational authority: behavior permitted, blocked, or changed: Advises subsequent prompt/memory revision. Behavioral-authority path: consumer, channel, force, horizon: Maintaining agent; self-assessment and observed effects; advisory; subsequent updates and sessions. Evidence source ID and local anchor: SRC-1, constitution/CONSTITUTION.md; passages see cited record. Claim IDs or none: RT-CLM-token-learning, EPI-CLM-general-learning. Mismatch marker or none: Outcome fit would not by itself validate a generalized explanation or transfer. Gap/limit: No candidate-linked evidence, oracle, result, blame assignment or subsequent consumer supplied.

Route ID: RT-RTE-memory-admission Route function: operational admission/selection/consumption Architectural status: doctrine only Object/candidate ID: EPI-OBJ-durable-learnings. Content/update relation: no content change. Transition or check target: File projection changes reaching underlying memory blocks. Evaluator/condition and domain: Successful remote push for file-based edits; external harness/Git synchronization condition, not semantic evaluation. Activation and timing: After commits and push; only file path covered. Possible or observed result: Propagation on success, no declared propagation without success. Implemented force: No inspected enforcement; declared permissive propagation gate. Epistemic authority and scope: None from push success; authorization/transport success does not certify truth. Operational authority: behavior permitted, blocked, or changed: File edits become underlying memory under the declared contract. Behavioral-authority path: consumer, channel, force, horizon: Synchronization machinery; pushed file changes; permissive gate; persistence into future inference. Exact implementation see record's limit. Evidence source ID and local anchor: SRC-1, constitution/AFFORDANCES.md; passage see cited record. Claim IDs or none: none. Mismatch marker or none: Operational admission is not evidence-consuming epistemic acceptance. Gap/limit: Direct memory tools and human edits are counterexamples to exclusive model authorship or a universal file gate; their semantic approval/veto rules are not disclosed.

Route ID: RT-RTE-memory-admission Route function: retention Architectural status: doctrine only Object/candidate ID: EPI-OBJ-durable-learnings. Content/update relation: no content change. Transition or check target: Proposed memory survives into underlying blocks/external memory and version history. Evaluator/condition and domain: Declared storage contract; agent or human edits; no truth judgment supplied. Activation and timing: Successful file propagation or unspecified bespoke memory write. Possible or observed result: Persisted candidate context, not demonstrated accepted knowledge. Implemented force: None inspected; declared durable availability. Epistemic authority and scope: Retention alone adds no warrant. Operational authority: behavior permitted, blocked, or changed: Makes candidate content available to later prompt construction or retrieval. Behavioral-authority path: consumer, channel, force, horizon: Future harness/model; memory blocks and external memory; supply, not mandatory compliance; across conversations and environments by contract. Evidence source ID and local anchor: SRC-1, constitution/AFFORDANCES.md; see cited record. Claim IDs or none: RT-CLM-token-learning. Mismatch marker or none: Retention is not lifecycle integration after acceptance. Gap/limit: No actual persisted candidate, trace-linked acceptance or exercised later reliance supplied.

Route ID: RT-BAP-system-prompt, RT-RTE-memory-admission Route function: behavior/policy adaptation Architectural status: doctrine only Object/candidate ID: EPI-OBJ-durable-learnings. Content/update relation: non-truth-apt policy/content update: changing persona, behavioral guidelines and future action cues; truth-apt assumptions are separately assessed above. Transition or check target: Updated system memory shaping later inference. Evaluator/condition and domain: Agent chooses guidance and external harness supplies it; compliance is model-interpreted. Activation and timing: Later inference after installation; exact packaging uninspected. Possible or observed result: Changed behavior, identity conditioning or correction against recurring mistakes; intended, not observed. Implemented force: None in repository; declared natural-language guidance, not a mechanical invariant. Epistemic authority and scope: No fact warrant follows from instruction priority. Operational authority: behavior permitted, blocked, or changed: Prompts can request different later action without first establishing a truth-apt proposition. Behavioral-authority path: consumer, channel, force, horizon: see RT-BAP-system-prompt. Evidence source ID and local anchor: SRC-1, constitution/CONSTITUTION.md, constitution/AFFORDANCES.md; see cited records. Claim IDs or none: RT-CLM-token-learning. Mismatch marker or none: Behavioral authority differs from epistemic warrant. Gap/limit: Delivery, activation and improved capacity are independently unestablished; no weight-update inference.

Route ID: RT-RTE-history-compaction Route function: content transformation Architectural status: doctrine only Object/candidate ID: EPI-OBJ-conversation-summary. Content/update relation: truth-apt transformation: indeterminate; advertised summarization aims at reshaping but no implementation or paired content establishes semantic preservation. Novel unsupported claims also cannot be inferred. Transition or check target: Prior messages to recursively summarized first message and references. Evaluator/condition and domain: Host size trigger; summarizer and fidelity evaluator unspecified. Activation and timing: Specified size exceeded; threshold unknown. Possible or observed result: Replacement summary supplied to later inference; raw messages remain retrievable by contract. Implemented force: None inspected; declared host-driven replacement and supply, not optional foreground selection. Epistemic authority and scope: Summary is a representation of conversation, not independently checked history or accepted conclusions. Operational authority: behavior permitted, blocked, or changed: Later inference operates on summarized rather than all original turns. Behavioral-authority path: consumer, channel, force, horizon: Later foreground model; first message of buffer; host-controlled supply, model-interpreted content; until further compaction or retrieval. Evidence source ID and local anchor: SRC-1, constitution/AFFORDANCES.md; see cited record. Claim IDs or none: RT-CLM-token-learning. Mismatch marker or none: Preserved raw storage does not prove summary fidelity or automatic correction. Gap/limit: No explicit rejection, fidelity comparison, acceptance or repair result supplied. The same record separately describes summary read-back/retention; no lifecycle integration after acceptance established.

Route ID: EPI-RTE-pruning-assessment Route function: check/evidence production Architectural status: doctrine only Object/candidate ID: EPI-OBJ-durable-learnings. Content/update relation: no content change. Transition or check target: Whether apparently situation-specific context can leave always-visible memory without losing necessary future guidance. Evaluator/condition and domain: Agent examines past conversations and judges retrieval readiness/confidence; domain is placement and future availability, not truth of the note. Activation and timing: Before externalizing or pruning context. Possible or observed result: Keep context, or judge external placement safe after index work; no observed result. Implemented force: None inspected; conservative policy. Epistemic authority and scope: Historical relevance/retrievability judgment only; not disconfirmation of the retained assertion. Operational authority: behavior permitted, blocked, or changed: Advises whether to proceed with externalization. Behavioral-authority path: consumer, channel, force, horizon: Maintaining agent; history inspection and retrieval-confidence judgment; advisory; next edit and future retrieval. Evidence source ID and local anchor: SRC-1, constitution/CONSTITUTION.md; passage see cited record. Claim IDs or none: EPI-CLM-general-learning. Mismatch marker or none: Removal from context is not epistemic rejection. Gap/limit: No quantitative confidence threshold, retrieval trial, decision record or guaranteed future retrieval.

Route ID: RT-RTE-memory-admission, RT-RTE-context-disclosure Route function: operational admission/selection/consumption Architectural status: doctrine only Object/candidate ID: EPI-OBJ-durable-learnings. Content/update relation: no content change. Transition or check target: Acting on pruning/externalization decisions changes which content remains always visible versus retrievable. Evaluator/condition and domain: Agent choice constrained by conservative identity and retrieval guidance; policy, not enforced rejection. Activation and timing: After placement assessment or task completion. Possible or observed result: Delete, compress, reorganize, or move content externally and update index; no observed withdrawal chain. Implemented force: None inspected; declared agent editing authority. Epistemic authority and scope: Changed accessibility does not retract a proposition's truth or establish its replacement. Operational authority: behavior permitted, blocked, or changed: Future model may cease receiving a note automatically and instead depend on index-triggered retrieval. Behavioral-authority path: consumer, channel, force, horizon: Future model; changed prompt/index and files; model-interpreted guidance; subsequent calls and tasks. Evidence source ID and local anchor: SRC-1, constitution/CONSTITUTION.md; EPI-RTE-pruning-assessment supplies the check passage. Claim IDs or none: RT-CLM-token-learning. Mismatch marker or none: No semantic rejection/rollback protocol established. Gap/limit: Raw-message retrieval and Git history are recovery resources, not evidence that forgotten context is noticed or restored.

Route ID: RT-RTE-subagent-context Route function: other — delegated recall and maintenance Architectural status: doctrine only Object/candidate ID: see cited record; memory-changing parts concern EPI-OBJ-durable-learnings. Content/update relation: truth-apt transformation: indeterminate; Recall imports history, Reflection may change claims or policies, Defragmentation targets organization. No concrete payload establishes equivalence, entailment or conjecture for their outputs. Transition or check target: Isolated conversation work or background maintenance to foreground information/shared memory. Evaluator/condition and domain: Foreground invokes specialist; internal checks, approval and exact results unspecified. No uniform truth evaluator across these roles. Activation and timing: Agent-invoked isolation/background work; no mandatory schedule. Possible or observed result: Recalled conversations, modified memory or reorganized memory, respectively; unobserved. Implemented force: None inspected; declared role-based capabilities. Epistemic authority and scope: Role names are not licenses for correctness or acceptance. Operational authority: behavior permitted, blocked, or changed: Shared memory may change outside the foreground message thread; isolation does not imply write isolation. Behavioral-authority path: consumer, channel, force, horizon: Foreground result channel uninspected; later model may receive shared-memory edits through prompt/retrieval; model-interpreted guidance; later conversations. Evidence source ID and local anchor: SRC-1, constitution/AFFORDANCES.md; see cited record. Claim IDs or none: RT-CLM-token-learning. Mismatch marker or none: Background reflection is not demonstrated acceptance or content-directed criticism. Gap/limit: No operative delegate evaluation result, veto, synchronization or merge consumer supplied; no automatic rejection/withdrawal semantics inferred.

No evidence-consuming disposition/acceptance decision or post-acceptance lifecycle integration is established in the inspected documents. This is a bounded evidence conclusion, not a claim that deployed Letta agents never accept claims. The documents explicitly allow verification-free updates while asking that their effects be observed; that countercase prevents treating every memory edit as accepted knowledge.

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-token-learning	see record	SRC-1, constitution/CONSTITUTION.md doctrine/design; README.md reported operation	Memory edits, read-back, self-assessment and prompt conditioning are described	None established; relevant declared routes are RT-RTE-memory-admission, EPI-RTE-update-assessment, RT-RTE-history-compaction, RT-RTE-subagent-context	None supplied; README use claim is not a trace	No interventions or outcomes. Changed context can condition action without proving correct knowledge or successful improvement	Explicit improvement-directed context policy, not demonstrated learning benefit	Behavioral adaptation, epistemic acceptance and improved capacity remain distinct.
EPI-CLM-general-learning	see record	SRC-1, constitution/CONSTITUTION.md; doctrine/design	Requests general rather than event-specific learning, historical checks and refinement	None established; declared EPI-RTE-update-assessment, EPI-RTE-pruning-assessment	None supplied	No tested environment/task contrasts; local success would not establish a proposed general mechanism	Generalization is an intended criterion for memory construction	No particular accepted ampliative claim, warranted derivation or transferred benefit established.

Bounded conclusion

The policy proposes acquired experience and retrieved context, durable memory edits, summaries, indexes and model-visible guidance. Acquisition makes information available but supplies no disclosed source-truth check; source warrant for actual retrieved/retained propositions is unknown. Summarization and generalization are declared, but actual transformation classes remain indeterminate without inputs, outputs or code. Neither an entailed derivation from warranted premises nor an accepted ampliative proposition can be identified from this source set.

Two distinct evaluative policies matter: memory-quality reflection asks whether prompts correct recurring mistakes and improve future user help; pre-pruning examination asks whether historical relevance and retrieval readiness justify removing always-visible content. Their intended uses and scopes are operational usefulness and placement, respectively. Neither supplies a recorded candidate-linked acceptance decision. Successful Git push is a separate propagation condition, not either evaluator. Thus retention and future operational use can precede or occur without established epistemic acceptance. No post-acceptance lifecycle integration can be traced.

Changed system memory has a declared later consumer, channel and horizon: the underlying model receives it as system-prompt context across calls; external context instead depends on relevance selection and retrieval. Both exercise natural-language influence rather than guaranteed compliance. Persona and guideline edits can directly adapt policy without producing a truth-apt claim. Compaction changes the model's evidence view without a supplied fidelity check; retrieving raw history is a recovery opportunity, not proof of repair. Reflection and Defragmentation can alter shared memory despite separate message contexts, but their result approval and withdrawal semantics remain unspecified.

The positive conclusion is an explicit improvement-directed context policy with retrospective assessment and later behavioral-authority paths by doctrine. Implementation, exercised reliance, acceptance, benefit and transfer are not shown. Distributed-parametric contents and change mechanisms remain uninspected. For the independent learning, reflection, theory-builder and self-improvement assessments, see RT-RTE-memory-admission: localized published theory does not establish consumption, content-directed criticism or iteration; guidance about self-representation does not demonstrate either causal direction in operation. No outcome or comparison closes the evidence-to-update-to-live-authority-to-changed-operation chain. These limits apply separately and do not negate the declared routes or imply a single system-wide epistemic verdict.

Shared records

Operative objects

EPI-OBJ-durable-learnings

Label: Durable learnings Part of: RT-RTE-memory-admission

This is the content-bearing candidate part of that supplied route, not another propagation mechanism. It consists of source-native corrected assumptions, user preferences, identity representations and refined behavioral guidelines. Its form is natural-language text, with claims and imperatives separately interpreted by the model. Declared storage is memory blocks projected to MemFS, including /system and external files; source-owned publication text is not an actual agent memory instance. SRC-1, constitution/CONSTITUTION.md, constitution/AFFORDANCES.md. The policy object RT-OBJ-context-constitution is distinct: it prescribes these updates rather than being their experience-derived content. Truth-apt scope depends on the particular user, environment and time, none of which is instantiated here.

As Letta agents accumulate experience, they should incorporate durable learnings (e.g., corrected assumptions, user preferences, and refined behavioral guidelines) and also build their own identities and sense-of-self in token-space representations. --- constitution/CONSTITUTION.md

• In-context memory: Memory files contained in /system represent in-context memories stored inside in the Letta agent’s system prompt (visible at all times) --- constitution/AFFORDANCES.md

EPI-OBJ-conversation-summary

Label: Conversation summary Part of: RT-RTE-history-compaction

This is the output text part of the supplied host-compaction route, not a new summarization implementation. It is natural-language summarized context plus references, stored as the first message of the in-context buffer under the declared harness contract. It differs from durable learnings in source material, trigger, producer and later delivery. No candidate instance or provenance chain is supplied. SRC-1, constitution/AFFORDANCES.md; minimum load-bearing passage see RT-RTE-history-compaction. Summary fidelity and actual truth-apt propositions are not determinable.

Routes

EPI-RTE-update-assessment

Label: Memory update assessment Part of: RT-RTE-memory-admission

This is the evaluative portion of the supplied experience-driven maintenance route, separated from generation, propagation and retention. It runs from experience and memory/prompt behavior to an agent's quality judgment, which is intended to guide later refinement. SRC-1, constitution/CONSTITUTION.md. The target is whether memory improves later help and corrects patterns of mistakes; the evaluator is the agent observing effects and asking self-questions. No actual judgment or decision is supplied.

• Immediate return: A proposed quality judgment, not a specified machine result or observed assessment.
• Later read-back: uninspected — observation and refinement over time are prescribed, but no assessment artifact is shown reaching a later invocation.
• Delegated visibility: uninspected — Reflection is mentioned elsewhere but no delivery of this assessment to a delegate is specified.
• Selection predicate: Memory quality, repeated mistakes, likely future user needs and effects of updates warrant reassessment.
• Invalidation or expiry: uninspected — no lifetime or supersession protocol for quality judgments is given.
• Activation or effect: Intended to refine memory and prompts over time; no operative result-to-edit chain supplied.
• Evidence limits: Policy questions and refinement instruction only; no evaluator implementation, oracle, candidate-linked test, result or acceptance decision.
• design conclusion status: claimed
• implementation conclusion status: uninspected
• operation conclusion status: uninspected

The agent proposes and judges changes in an open-ended ongoing mode. Human preferences are assessment subject matter; human per-change approval is not established. There is no supplied expected-answer provider. Guidance favors durable, incremental and generalizable learning. The parent route supplies the declared write gate and persistence; no independent semantic veto or rollback mechanism for this assessment is specified. It can motivate further editing, but it does not itself demonstrate changed retained theory or evidence-consuming acceptance.

• If I encounter a similar situation in the future, will I be able to better meet my user’s needs?
• What are the patterns of mistakes that I am making, and do my prompts correct against those mistakes?
• Will the user be disappointed if I don’t remember this information in the future? --- constitution/CONSTITUTION.md

The qualification that many token-space updates lack explicit reward or verification, followed by the instruction to observe and refine effects, is quoted in RT-RTE-memory-admission. That passage is a source-supported counterexample to mandatory pre-admission verification. The parent record's separate theory-builder conditions 1–4 and learning/reflection/self-improvement limits apply; no new criticism/revision episode is established by these questions.

EPI-RTE-pruning-assessment

Label: Pruning placement assessment Part of: RT-RTE-memory-admission

This is the pre-removal evaluative part of the parent's edit/delete/reorganize path. It is distinct from the preceding quality assessment: the target is safe placement and future retrieval, not whether an update improves user help. It proceeds from a proposed externalization through examining past conversations to an agent confidence judgment about the index and later retrieval. SRC-1, constitution/CONSTITUTION.md.

• Immediate return: Judgment to retain in-context guidance or consider it safe to move externally; no specified result schema.
• Later read-back: uninspected — future retrieval through an updated index is intended, but no actual decision/index/read chain is supplied.
• Delegated visibility: uninspected — no delegated review or transmission of this judgment is specified.
• Selection predicate: A note or correction appears unnecessary in every conversational turn, or pruning is contemplated.
• Invalidation or expiry: uninspected — confidence revision and stale-index invalidation are not specified.
• Activation or effect: Intended to constrain externalization until reliable retrieval groundwork exists; compliance and effect unobserved.
• Evidence limits: Natural-language policy with no confidence threshold, executable rejection, retrieval test, actual result or recovery trace.
• design conclusion status: claimed
• implementation conclusion status: uninspected
• operation conclusion status: uninspected

The agent proposes removal and evaluates historical relevance/retrieval readiness. No separate human approval or veto is required by this passage; human editing remains possible under the parent route. This is ongoing context maintenance, not a bounded experiment with an answer oracle. A favorable judgment licenses a placement change by policy; no implemented enforcement or semantic truth rejection follows. The parent's declared memory-write pathway governs persistence, while retrieval and Git history remain resources rather than automatic rollback.

Similarly, Letta agents should avoid making assumptions about what context is necessary to have in every conversational turn. Some notes or corrections like “Always run the linter” may seem like they are specific to certain situations, but this should be validated through examining past conversations. Context should only be removed from the in-context memory and placed externally if the Letta agent is confident that it has laid the groundwork (e.g. through updating its index) to retrieve the context reliably when necessary in the future. --- constitution/CONSTITUTION.md

Claims

EPI-CLM-general-learning

Label: General learning requirement

The Constitution asks that learned context improve future action across changing conditions instead of simply remembering each event. This is a normative transfer requirement and proposed benefit, not an observed generalization result. SRC-1, constitution/CONSTITUTION.md. It is a narrower, distinct claim within the broad token-space improvement objective RT-CLM-token-learning, not a second record for the same wording. Assessment requires evidence about actual context changes and later situations, not just the existence of memory or complete logs.

Learning should be done in a general way. Letta agents should not expect future situations to only repeat exact past circumstances. They will move across new environments, work across different projects, and run across different models for a variety of tasks: the context they curate should therefore help them improve in the future, regardless of changing conditions, rather than simply creating memories for every experienced event — recall that Letta agents automatically store all prior experiences in a queryable database. --- constitution/CONSTITUTION.md