Context Constitution synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
Context Constitution supplies an explicit improvement-directed policy: turn experience into durable corrections and guidance, then use that context in future work. Its contribution is the specification of that loop, not demonstrated learning benefit. This account covers the complete published artifact at commit 964f30137ace9eade1aef112992ed6539b61bbb0, with an applicability cutoff of 2026-10-15. It does not cover an executing Letta harness. The boundary is complete artifact, partial loop, and the evidence is doc-grounded. SRC-1 supplies policy and affordance descriptions in constitution/CONSTITUTION.md and constitution/AFFORDANCES.md, plus reported internal prompting/training use in README.md. The only shipped executable check concerns Markdown whitespace. No agent execution or causal experiment is supplied; see the boundary and source register and RT-CLM-token-learning.
Establish a continuing context, then disclose detail
The declared runtime places a continuing agent between a user and an underlying language model. The external harness supplies the system preamble, tool interfaces and conversation buffer. The foreground model chooses retrieval and editing actions; the host owns automatic message storage and size-triggered compaction. The documents do not implement model-call scheduling, tool dispatch, permissions or task termination. The underlying model is an external distributed-parametric component whose exact version is not pinned. User model replacement is permitted; continuity is assigned to retained context, not invariant weights. This does not establish continuity across an actual replacement or any online parameter update (RT-CMP-underlying-model).
Memory placement determines later availability. Underlying memory blocks are the declared authoritative store; MemFS is a Git-tracked local projection. /system prose supplies standing identity, goals and guidance on subsequent inference calls. External memory, including procedural skills, instead requires an explicit read. This is model-interpreted influence, not enforced compliance (MEM-OBJ-memory-blocks, MEM-OBJ-mem-fs, MEM-OBJ-external-memory, RT-BAP-system-prompt).
The representative workflow follows an ordered source example: recognize a task situation from an in-context cue, inspect the referenced skill's metadata, and load the full skill only if useful. Full content can lead to further references. The foreground agent owns each decision; a negative usefulness judgment can stop expansion. Automatic skill/tree/history metadata and agent-maintained indexes provide discovery cues before the requested read. Baseline supply and requested full-content delivery are therefore separate operations. Retrieval establishes access, not source endorsement or a correct task result (MEM-OBJ-context-metadata, RT-RTE-context-disclosure, workflow account).
For a task needing occasional detailed procedures, this design proposes a compact baseline with conditional expansion. Its adequacy depends on useful cues and successful later retrieval. The policy asks agents to release completed-task material, but exposes no precise selective-eviction primitive or numerical budget. It also forbids treating efficiency as unconditional token minimization: identity preservation and confidence in future retrieval constrain removal. The supported workflow ends at usable-context supply and prescribed cleanup, not verified delivery to a user. No concrete request, answer oracle, task acceptance rule or successful delivery trace is available.
Convert experience into later guidance
Durable maintenance is optional follow-up, not a mandatory stage after every retrieval. The agent should generalize corrected assumptions, preferences and behavioral guidance rather than copy every event. Agent-selected updates and explicit human file edits are both described. For the file route, local editing and committing are insufficient: successful remote push is the declared propagation gate into underlying blocks. That gate is an external synchronization contract, not an inspected enforcement mechanism or a semantic approval test. Bespoke memory tools and delegated writes have unresolved admission rules; the file gate cannot be generalized to all writes (RT-RTE-memory-admission, EPI-OBJ-durable-learnings).
Two evaluative policies qualify maintenance. Memory-quality assessment asks whether prompts correct recurring mistakes and improve future help. Pre-pruning assessment asks whether historical relevance and an adequate index justify moving guidance out of always-visible memory. These address usefulness and placement respectively, not a universal truth criterion. The policy acknowledges that updates often lack explicit reward or verification and requests observation and refinement over time. No candidate-linked, evidence-consuming acceptance decision is supplied. Retention and operational use consequently cannot be reported as accepted knowledge or post-acceptance integration (EPI-RTE-update-assessment, EPI-RTE-pruning-assessment).
Actual truth-apt transformation remains indeterminate without concrete inputs and outputs. A generalized memory could import, reshape, derive or conjecture content; the instruction to generalize establishes none of those relations for a particular claim. Persona and guideline edits can instead change policy without making a truth-apt assertion. The intended downstream authority is clear, but actual delivery, meaning-dependent activation and benefit remain separate unresolved questions; see the epistemic conclusion.
Continue through compaction and delegated maintenance
Raw message history is declared automatically retained and searchable, including tool calls and returns. When a specified context size is exceeded, the host supplies a summary as the first buffer message, preferentially replacing older in-context history. Recursive summaries and references carry prior context into later inference. This is a declared automatic trace-fed update with a later consumer, not merely raw logging. It supports a claimed trace-learning pathway without establishing new knowledge, semantic fidelity or improved capacity. Raw-history search provides a recovery opportunity, not proof that omissions are detected or repaired. The desired extraction of key facts into external memory has no specified automatic checkpoint producer or admission route (MEM-OBJ-message-history, MEM-OBJ-summary-message, RT-RTE-history-compaction).
Delegation is an auxiliary affordance, not a required stage of the representative workflow. Recall retrieves conversations; Reflection reviews conversations to modify memory; Defragmentation reorganizes memory. Separate message threads can keep intermediate work out of the foreground window, but the same agent's conversations share persistent memory. Isolation therefore does not imply return-only effects, write isolation or a security boundary. Handoff inputs, return payloads, approval, concurrent merge ownership and failure recovery are unspecified (RT-RTE-subagent-context).
What the improvement account establishes
The strongest supported conclusion is an explicit policy for experience-responsive context revision, retrospective assessment and future model-visible guidance. It also requests generalization across projects, environments and models rather than event-specific recall. No comparison establishes that these prescriptions improve future capacity or transfer (EPI-CLM-general-learning). The independent property assessments in RT-RTE-memory-admission remain distinct:
- Theory-builder condition 1, localized content: claimed. The published policy states inspectable claims about context, identity and learning. Condition 2, consumption; condition 3, content-directed criticism; and condition 4, iteration: uninspected, each separately. No traced decision depends on the policy's meaning, and no challenged claim, blame assignment, revision and next round is supplied.
- Learning through criticism: no episode establishes that criticism of a consumed theory improved future capacity. This does not negate the proposed trace-fed context updates.
- Reflection: self-representation and reciprocal influence between identity and experience are prescribed; both causal directions in operation remain uninspected. A Reflection role name does not demonstrate the property.
- Autonomy: autonomous theory building remains unestablished. Model proposal authority does not show computation performing every internal operation, particularly with unresolved admission and coordination.
- Self-improvement: the assessment boundary is the agent's context-determining organization across later sessions, with independently stated objectives of better future help, fewer repeated mistakes and better retrieval. The update pathway is claimed. No trace closes the links from objective-bearing evidence to a particular update, admission into that organization, live behavioral authority for that exact update, and subsequent operation causally depending on it. Exercised self-improvement and successful improvement therefore cannot be attributed separately.
A deployed prompt and synchronization trace would strengthen implementation and installation findings. A retained correction episode linking evidence, decision, admitted guidance and a later meaning-dependent action would change the consumption, criticism, iteration and exercised-change assessments. Matched outcome comparisons, with the relevant component independently varied when component attribution is sought, would be needed to attribute benefit. Paired pre/post-compaction buffers and checks against raw history would change the fidelity assessment. None of these missing observations licenses an absence claim about external Letta deployments. Both the report verification and profile verification accept the bounded account with no additional declared limits; they do not upgrade its evidence strength.
Limitations
| limitation | affected SRC IDs and record citations | inspected boundary | conclusion prevented | evidence that would resolve it |
|---|---|---|---|---|
| Policy artifact without its executing integration | SRC-1; RT-OBJ-context-constitution, RT-BAP-system-prompt, RT-CMP-underlying-model | Complete frozen document tree; harness, deployed prompts and internal training are not shipped | Actual policy loading, model compliance, cross-model continuity or deployed reliability | Frozen consuming integration, exact deployed inputs and execution traces, including model changes where relevant |
| Partial write/admission knowledge | SRC-1; RT-RTE-memory-admission, MEM-OBJ-memory-blocks, MEM-OBJ-skill-assets | Declared file push contract; unspecified initialization, backend, direct-tool and delegated gates; opaque asset consumers | Universal push or human-approval requirement, semantic validation, complete storage/form inventory, executable skill authority or rollback guarantees | Admission and synchronization implementations, initialized state/provenance, actual asset payloads and consumers, and failure/recovery traces |
| Discovery and delivery do not establish task success | SRC-1; RT-RTE-context-disclosure, MEM-OBJ-context-metadata | Prescribed cue–metadata–full-content workflow; no concrete request-to-result run | Reliable retrieval, source endorsement, measured context savings, implemented selective eviction or verified user delivery | Selector and budget implementation, read/cleanup traces, named task acceptance criteria and evaluated delivered results |
| Summary fidelity and recovery are unresolved | SRC-1; RT-RTE-history-compaction, MEM-OBJ-message-history, MEM-OBJ-summary-message | Declared threshold compaction and searchable raw history; no summarizer, paired buffers or repair episode | Losslessness, exact retained tail, rationale preservation, prior-summary persistence, automatic fact checkpoint or successful recovery | Transformation implementation, paired source/summary artifacts, checkpoint wiring, fidelity decisions and retrieval/repair traces |
| Delegation lacks handoff and coordination evidence | SRC-1; RT-RTE-subagent-context | Named specialist roles and isolated message threads with shared memory | Return-only effects, known permission/merge rules, failed-job recovery or successful delegated delivery | Delegate inputs, grants, writes, merge/rejection decisions, completion returns and failure-handling traces |
| No candidate-linked acceptance, criticism episode or improvement comparison | SRC-1; RT-RTE-memory-admission, EPI-RTE-update-assessment, EPI-RTE-pruning-assessment, EPI-CLM-general-learning | Published generalization and assessment policies; no retained candidate instance or observed outcome | Accepted new knowledge, checked derivation, content-directed criticism improving capacity, exercised reflection/autonomous theory building/self-improvement, causal benefit or transfer | Candidate provenance and input/output content; evidence-consuming decisions with criterion, intended use and scope; retained criticism/revision and subsequent consumer traces; appropriately controlled outcome comparisons |