Dense-Mem synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From caller assertions to retained memory
Dense-Mem is a self-hosted memory service, not the conversational controller of an enclosing agent. It claims responsibility for durable evidence, ownership, lifecycle, support eligibility and bounded recall; the host owns conversation and judgment. This analysis covers the shipped server, clients, portals, intake, maintenance, prompts and evaluators at commit bd736450454e237f1ab6d138f4ae9755916d73f6, with cutoff 2026-10-15. Its single source, SRC-1, supplies inspected implementation and target-authored doctrine, not observed execution or causal experiments. Whole-system inclusion does not mean exhaustive inspection. External providers, infrastructure and unrepresented host behavior remain outside the source boundary (boundary, Source register).
The production driver composes PostgreSQL repositories, authenticated MCP/browser entry points and background workers. Code owns progression. Remote models propose semantic interpretations, extraction and organization; embeddings supply numerical search representations. Model roles are separately configured, their exact weights are unpinned, and the inspected calls do not update their parameters. A configured name can resolve through a mutable endpoint (RT-CMP-openai-assessor, RT-CMP-openai-embedding, RT-CMP-openai-verifier).
An ordinary Remember request supplies evidence, relationship proposals and an idempotency key. Authentication, team/scope controls, input/security checks, server-selected candidates, bounded assessment/repair and repository fences constrain admission. The assessor judges proposals against supplied evidence; software checks authorized references, exact citation ranges, response coverage and preserved tuple/temporal fields. Accepted relationships acquire evidence-span and assessment lineage. This is an evidence-consuming acceptance criterion for semantic storage, not independent authentication of source truth or proof of faithful interpretation. Safe evidence can be retained when every relationship is declined (RT-RTE-remember-intake, EPI-OBJ-relationship-proposals, EPI-OBJ-evidence-content).
Direct Remember commits accepted evidence, semantic state, search projections and a terminal result after provider work. Transactional replay/duplicate and active-contract fences govern accepted writes. Completed or non-retryable attempts replay their stored result; changed content under the same key conflicts, while a later eligible call can retry a retryable failure. Optional outer locking is not an exactly-once provider-call guarantee. These controls depend on PostgreSQL transaction/locking and trusted scope contracts, and do not establish deployed durability (RT-CLM-commit-fences).
Private session intake is materially different: raw events are staged before extraction, and validated extraction/linking checkpoints can survive even when semantic admission fails. Retry reloads them rather than reconstructing a host conversation. New intake automatically supplies selected completed same-session events, at most ten and within 2048 prior-context tokens, to bounded extraction frames. Linking receives proposals and cited segments; evidence commit checks exact staged-event substrings. The retained chain is raw events → extraction/linking checkpoints → semantic preparation → evidence/relationships → later intake or recall. Exact spans establish provenance, not universal entailment (RT-RTE-session-intake, MEM-OBJ-session-events, MEM-OBJ-extraction-checkpoints).
Context control is local to these tasks. Remember may substitute an oversized assistant response with a fixed placeholder while preserving the request and validation feedback; this is omission, not a durable summary or reload pointer. Session repair regenerates from base messages and new feedback without accumulating invalid responses. Selected inputs enter a bounded model frame and parsed outputs escape; host compaction and post-recall prompt assembly are not established by these mechanisms (RT-OBJ-assessment-session, RT-RTE-session-intake).
Maintaining reliance and organization
Retained evidence prose, symbolic semantic/support history, searchable text and dense vectors have different consumers and authority. Text supplies context; vectors and text influence retrieval ranking; ownership, generation, lifecycle and support state constrain software eligibility. PostgreSQL persistence is distinct from vector access, and vectors are retained outputs rather than updated provider weights (MEM-OBJ-evidence-fragments, MEM-OBJ-relationship-records, MEM-OBJ-search-documents, MEM-OBJ-search-vectors).
Later changes do not share one admission gate. Automatic ontology maintenance publishes derived classification/equivalence under software policy, with deterministic handling of locked decisions and bounded model proposals otherwise. Manager publication instead exposes explicit draft, publish/discard and rollback under role, grant, audit and revision checks. A manager role does not prove a human actor, and automatic publication does not require per-item manager approval. Current definitions and assignments feed later organizer context and recall grouping; receipts retain bounded outcomes and reasons, not necessarily full deliberation. Organization changes discovery, not factual evidence (RT-RTE-ontology-maintenance, RT-RTE-manager-publication, MEM-OBJ-ontology-records, MEM-OBJ-organization-receipt).
Conflict review can select a dossier position through deterministic policy or a verifier, subject to position validity and confidence of at least 0.70. Abstention or sufficient failures can instead produce last-write-wins fallback; rejecting a model answer therefore need not mean no change. Its deletion-only derived fragment is audit/recovery material, excluded from ordinary evidence recall and unable to establish a semantic relationship. Changed relationship eligibility and retained audit text must not be conflated (RT-RTE-conflict-review, MEM-OBJ-conflict-derivation).
Owner correction can supersede a tuple and copy effective supports under version, ownership and support-set checks. Its server-policy entailed label does not establish a fresh semantic entailment check of the patched proposition. Retraction revokes support and recomputes eligibility; independent support can keep a relationship active. Private erasure instead seals and advances a space generation before fenced manifest deletion and audit scrubbing. A queued operation is not completed erasure, and external copies are outside its guarantee (EPI-RTE-relationship-correction, EPI-RTE-evidence-withdrawal, MEM-RTE-private-erasure).
Returning context and reaching delivery
Recall is requested read-back, not unsolicited insertion into a host conversation. It selects under query, temporal, space, identity and result-count bounds, with lexical fallback possible when vectors are unavailable. Internal automatic supply also exists: prior-session context and organizer vocabulary are selected for their next consumers. These internal selections do not demonstrate host prompt assembly (RT-RTE-recall-context, MEM-OBJ-session-events, MEM-BAP-organization-selection).
A consequential limit is stronger than missing execution evidence: the README's claimed mandatory final relationship-support gate is not enforced in the inspected ordinary lexical path. With organization disabled, an otherwise visible query-matching evidence document can survive optional relationship attachments and final selection with an empty relationship array. This is a conditional implementation counterexample, independently upheld by report verification, not an observed deployment incident or a finding about every vector/organized branch. Other lifecycle, quarantine, source and space restrictions still apply. Returned context is source-fragment text subject to truncation, not necessarily the full fragment (EPI-CLM-support-gating, EPI-RTE-recall-hydration, report verification).
Community discovery carries advisory guidance to verify details with trace before using them as support. Trace exposes bounded lineage but neither forces a caller's acceptance decision nor guarantees complete reconstruction. Pack export serializes selected active relationships and optional provenance into hashed JSON; a hash identifies bytes, not truth. The separate skill-export prompt affords host drafting and review, expressly not automatic memory import or established skill installation (RT-BAP-discovery-guidance, EPI-RTE-trace-lineage, EPI-RTE-pack-export, RT-RTE-skill-export).
The representative shipped workflow makes the delivery boundary concrete. The Python example submits Remember, repeats it, asserts response equality, requests recall, traces the original Remember splits and returns the original result. It discards recall and trace outputs; it does not verify every recalled discovery or assess trace semantics. SDK errors stop progression, cancellation leaves receipt unknown, and a later recall/trace failure does not roll back an earlier Remember commit. This source-defined sequence establishes neither successful execution nor answer benefit (WF-CLM-client-delivery).
Capacity and assessment-changing evidence
The strongest supported experience-related contribution is a wired conditional trace-fed update: when session text records interaction experience, automatic extraction creates durable material with later consumers. Actual input origin remains mixed or opaque, so neither every session input nor a specific trace-source category is established. This durable behavior-shaping route does not demonstrate improved capacity for future action. Feedback retains attributed caller judgments; evaluation derives retrieval metrics from traces, mappings and qrels with optional gates. Neither inspected path establishes automatic installation of an improved policy. No achieved score or component effect is supplied (MEM-OBJ-extraction-checkpoints, MEM-OBJ-session-events, MEM-RTE-feedback-recording, RT-RTE-evaluation-suite, profile verification).
These contributions do not resolve the independent theory and improvement questions. The accepted runtime account establishes bounded semantic judgment, maintenance and admission, but does not establish theory-builder conditions 1–4 for a formulated theory route: localized theory content, content-dependent consumption, content-directed criticism and retained iteration are not separately demonstrated as such. Validation repair alone does not establish criticism of a theory. Criticism of a consumed theory improving future capacity is unestablished. Automatic maintenance progression is wired; autonomous theory building is not established. Reflection about the bounded service is also unestablished, rather than disproved by opaque models or retired Dream displays. Exercised self-improvement is unestablished: retained organization can reach later consumers, but evidence-responsive change directed at an independently specified improvement objective and subsequent operation causally depending on that change are not demonstrated. Successful improvement requires additional outcome evidence, independently of these membership questions (RT-RTE-remember-intake, RT-RTE-ontology-maintenance, RT-RTE-evaluation-suite, MEM-RTE-administrative-views, epistemic conclusion).
For evidence-linked storage, recovery and selective context delivery, the records establish concrete implementation mechanisms, not deployment reliability. For a use requiring every evidence hit to carry active relationship support, the inspected ordinary path does not provide the claimed guarantee. For end-to-end answers or improvement, the missing host/run/comparison evidence prevents assessment. A shared-path support predicate or a narrowed public contract would change the gate assessment. Pinned provider outputs and source-to-proposal comparisons would test semantic faithfulness; scoped execution artifacts would establish activation; controlled interventions with downstream outcomes would support benefit attribution at the level actually varied. Complete transaction and rollback inspection would narrow freshness/reversal uncertainty. These are assessment-changing evidence or system changes, not a product ranking.
Limitations
| limitation | affected SRC IDs and record citations | inspected boundary | conclusion prevented | evidence that would resolve it |
|---|---|---|---|---|
| Reconciliation retained a text-delivery conflict; report verification dismisses a completeness reading but preserves truncation. | SRC-1; RT-RTE-recall-context, EPI-RTE-recall-hydration, report verification | Memory's source-content wording versus ordinary hydration's bounded context | Delivery of the entire original fragment; the settled source-lineage interpretation does not license completeness | A full-content route contract and producer-to-consumer evidence for any stronger delivery claim |
| No supplied target execution, provider calls or intervention artifacts. | SRC-1; RT-RTE-remember-intake, RT-RTE-evaluation-suite, WF-CLM-client-delivery | Source wiring and shipped example/evaluator, not deployment | Exercised reliability, achieved scores, host reliance or causal benefit | Inspectable scoped runs and controlled comparisons with stated outcomes |
| The support-gating counterexample covers ordinary lexical recall with organization disabled. | SRC-1; EPI-CLM-support-gating, EPI-RTE-recall-hydration | Candidate, hydration and final selection in that path | Universal gating guarantee, observed incident, or identical behavior of every recall branch | Branch-specific tracing and execution evidence; a common enforced gate for a universal guarantee |
| Semantic faithfulness, duplicate identity and complete historical consistency are not audited. | SRC-1; EPI-OBJ-relationship-proposals, MEM-OBJ-evidence-fragments, MEM-OBJ-extraction-checkpoints | Contract/span checks and bounded intake/hydration reads | Source truth, universal entailment, perfect extraction/canonicalization or temporal consistency | Concrete source/output comparisons and complete relevant validator, canonicalization and history-path inspection |
| Full transactional freshness, projection completion and reversal rules remain uninspected. | SRC-1; MEM-OBJ-ontology-records, MEM-RTE-derived-projections, RT-RTE-manager-publication, MEM-RTE-lifecycle-changes | Service/current-state checks and selected repository paths | Universal stale-write prevention or generic undo | Complete transaction/rollback tracing and scoped race/reversal evidence |
| Historical payloads, initialization, full diagnostics and rationale consumers remain unresolved. | SRC-1; MEM-RTE-administrative-views, MEM-OBJ-organization-receipt, RT-OBJ-assessment-session | Display summaries, receipt contracts and bounded working history | Complete memory inventory, full rationale reuse, or reflection inferred from historical names | Payload/provenance/admission inspection and actual later-consumer tracing |
| Caller roles do not authenticate human control or original trace content. | SRC-1; MEM-OBJ-session-events, RT-RTE-manager-publication, MEM-RTE-lifecycle-changes | Generic session/framework, owner and manager identities | All-session trace learning, a specific original trace-source classification, or required human approval | Content-origin evidence and explicit per-content operator-decision records |
| Provider parameters, deployment contracts and external host behavior are outside available evidence. | SRC-1; RT-CMP-openai-assessor, RT-CMP-openai-embedding, RT-CMP-openai-verifier, RT-RTE-recall-context | Configured transport/model roles and service returns | Immutable model identity, provider training behavior, deployed isolation/durability or enforced host compliance | Pinned provider identity/contracts, deployment evidence and host-consumption traces |
| Erasure, display and export have narrower guarantees than total deletion or complete portable provenance. | SRC-1; MEM-RTE-private-erasure, MEM-RTE-trace-export, MEM-RTE-feedback-recording | Manifest transaction, bounded trace/export and optional feedback composition | Verified erasure of external copies, complete UI/SQL authorization audit, or lossless provenance delivery | Full manifest/retention and authorization inspection, completion artifacts, recipient/copy accounting and explicit completeness checks |
| Broader feedback/CI consumers, evaluation import/baseline provenance and skill adoption are not established. | SRC-1; MEM-RTE-feedback-recording, RT-RTE-evaluation-suite, RT-RTE-skill-export | Recording/readers, scoring dispatch and host prompt | Complete absence of adaptation, reproduced benchmark quality, release-wide enforcement or installed skill authority | Consumer/automation and artifact-provenance inspection; traced review, installation and subsequent use |
Both supplied verifications declare no additional limits or blockers. The delivery disagreement above is retained with its verification disposition, not promoted into an unresolved behavioral conflict. This synthesis uses accepted members only and adds no source reads or evidence.