SAGE synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From external agent work to retained memory
SAGE is a memory service and coordination layer for external agent runtimes, not an enclosing conversational executor. It retains project decisions, observations and lessons, then supplies selected material to later agent work. This account covers the shipped system at commit e2e0242a421afbd356ed135d7d754a35b0ad5f14, with analysis cutoff 2026-10-15. Its single registered source, SRC-1, is the frozen l33tdawg/sage repository. Evidence is inspected implementation and source-native doctrine, with attributed research reporting distinguished below. No target execution, model response or causal experiment was established. Whole-system inclusion does not mean every included branch was inspected. See boundary and coverage and the Source register.
An external model decides its next action and proposes what to remember. SAGE's personal node or daemon owns signed service requests, admission, persistence, retrieval and background maintenance. In the ordinary MCP sage_turn path, service code recalls by topic and domain, optionally submits an observation, checks incoming work and returns control to the model. Recall and storage can fail independently; the combined call is not an atomic transaction. A successful submission transaction initially leaves the memory proposed, not necessarily accepted or immediately retrievable. RT-RTE-turn-cycle, RT-CLM-atomic-turn.
Retained prose, canonical envelopes and retrieval indexes are different operative parts. Canonical state uses Badger; serving content uses SQLite on the personal node and PostgreSQL on the daemon. Text is interpreted by agents, while symbolic access, lineage and lifecycle fields govern service decisions. Dense vectors support retrieval rather than constituting the recalled advice. Co-commit can retain a hash-only envelope with readable content staged off-chain, so canonical inclusion alone does not imply another node can reconstruct the prose. MEM-OBJ-memory-record, MEM-OBJ-memory-indexes.
Admission has several meanings
Ordinary submissions undergo authority and transaction checks, a node's quality vote and global or scoped quorum. Signature and replay checks are repeated in consensus execution under activated version rules, rather than relying only on front-end screening. Baseline suitability checks cover duplication, length/noise, confidence and domain, not factual truth; a dedup lookup error passes that local check. A single-node quorum is not independent corroboration. Persistence failure handling can halt rather than report successful Commit after a known flush failure, but this is not verification of every crash state. RT-RTE-memory-admission, EPI-RTE-baseline-check.
The strongest supported semantic contribution is an optional, consequential evidence-support route. Configured judges distinguish lasting knowledge from session remarks and, when support judges and submitted evidence are available, assess whether that evidence supports the assertion as stated, including time and certainty. Lastingness and support are separate checks; both must pass when both apply. Missing judgments hold the in-scope node vote, and explicit evidence expiry holds rather than silently becoming an evidence-free submission. This is implemented evidence-consuming acceptance for later memory use within the configured node/domain scope. It neither authenticates the evidence nor certifies external truth. EPI-RTE-lasting-check, EPI-RTE-support-check, EPI-RTE-semantic-disposition.
An operator can resolve a semantic hold. That retained accept/reject decision survives judge-version changes and is read before the model verdict, but only after fresh baseline and scope checks. The voter consumes its enum, not its optional rationale note; operator acceptance cannot rescue baseline failure or directly commit the memory. Supporting evidence is temporary and can be deleted after terminal disposition even though the memory and local verdict outlive it. Later recalling agents are not established to receive this dossier. MEM-OBJ-review-decision, MEM-OBJ-gate-material, EPI-RTE-evidence-retention.
These guarantees are route-specific. Coauthor-ratified co-commit becomes committed through block inclusion and bypasses the ordinary semantic voter. Federated copies instead become locally authorized proposals subject to receiver voting. Current dashboard imports use consensus admission, while historical branches can insert serving rows directly. Consequently neither “committed” nor “imported” means universally semantically checked. RT-RTE-co-commit, MEM-RTE-sync-copy, MEM-RTE-import-material.
Later use depends on selection and the receiving host
Requested recall combines caller authorization, lifecycle, confidence, search mode, vector-space compatibility and optional reranking. Challenged-but-live records can remain eligible with reduced confidence: MCP preserves dispute warnings, whereas the browser turn omits status/dispute fields. Thus relevance and delivery qualification differ by client, and a returned result is neither endorsement nor complete store visibility. RT-RTE-targeted-recall.
Host hooks provide a separate context route. SessionStart pushes up to ten recent domain-scoped excerpts, each truncated to 200 bytes. Default-off, consent-gated pre-compaction capture selects user/assistant text, excludes tool bodies and metadata, encodes chunks and tracks admission/retry progress. Resumed-thread recall matches durable session/thread identity and reconstructs eligible committed units. Capture must not block host compaction; errors or deadlines can prevent reconstruction output. SAGE reloads text by value but neither owns context eviction nor guarantees that all stdout fits the host's model window. This is selective external retention and reload, not full transcript preservation or prevention of compaction. RT-RTE-hook-context, MEM-OBJ-capture-unit, MEM-OBJ-capture-progress.
Inception also reloads operator boot text and mode-dependent operating instructions. Full, bookend and on-demand modes differ from the shipped skill's blanket per-turn prescription. Saved custom text supplies directions through a tool response, not a privileged system message or enforced execution policy. The external client remains responsible for interpreting recalled content and checking current authorization. RT-RTE-boot-guidance, RT-BAP-client-guidance.
Revision, feedback and future capacity
SAGE implements two particularly clear experience-to-later-operation pathways. Recorded votes and the first terminal memory verdict update durable validator statistics that later quorum calculations consume as weights. This is numerical adaptation, with consensus agreement as its reference—not independent factual ground truth. Separately, the Level Up bridge feeds bounded challenge/outcome data and selected prior lessons to a summary model, submits the resulting observation and makes admitted outputs available to later query and summarization calls. These are wired trace-fed updates, not merely log storage. Whether the bridge compresses, derives or conjectures cannot be settled without full input/output pairs; requested novelty establishes no particular semantic transformation. Neither pathway establishes useful learning. MEM-RTE-verdict-feedback, MEM-OBJ-validator-stats, RT-RTE-bridge-distillation.
Other revision surfaces preserve important distinctions. Reflection storage formats caller-authored summary/DO/DON'T text into proposals; assigning a DO the fact type and 0.90 confidence does not prove success or promote an already retained entry. Correction admits a replacement and independently reads back its committed status before attempting to challenge the original. Interruption can leave both live; the check protects availability, not truth. Challenges, restricted reinstatement, decay and operator-authorized cleanup change reliance or eligibility without necessarily deleting bytes or establishing falsity. Re-embedding changes search representation, not prose or provider weights. MEM-RTE-reflection-storage, EPI-CLM-reflect-fact, MEM-RTE-correction-cycle, MEM-RTE-lifecycle-change, RT-RTE-maintenance-cleanup, MEM-RTE-reembed-index.
For actual bridge outputs, theory-builder condition 1, a whole formulated theory, remains uninspected; condition 2, decisions depending on its content, remains uninspected; condition 3, blame-bearing criticism and revision or changed reliance, remains uninspected; and condition 4, retained criticism shaping another round, remains uninspected. Record-level addressability and cross-session persistence are supported capacities, not substitutes for those independent conditions. Whether criticism of a consumed theory improved future capacity is likewise uninspected. Reflection as a causal self-representation and autonomous theory building are independently unestablished; an operation named reflect or a prompt requesting criticism establishes neither. These assessments concern the traced routes and supplied outputs, not a system-complete absence. RT-RTE-bridge-distillation, MEM-RTE-reflection-storage, MEM-RTE-correction-cycle.
For the frozen SAGE service plus shipped consumers, across later admissions and sessions, feedback can durably change decision inputs and supplied advice. Exercised self-improvement is not established: no subsequent operation is supplied that causally depends on those updates toward an independently specified improvement objective. Numerical agreement feedback does not itself show improved correctness. The source reports longitudinal learning in a memory-backed organization, but discloses that underlying pipeline and run records are excluded; this analysis preserves the report without independently validating its statistics or attributing an effect to consensus or judging. Fixed inference weights would not exclude learning through retained advice, but memory writes alone do not establish learning or model training. MEM-RTE-verdict-feedback, RT-BAP-retained-context, EPI-CLM-feedback-learning.
A representative coordination workflow
The exact-local Messages protocol supplies a coherent multi-step workflow without inventing a SAGE-owned worker. An external sender supplies an exact recipient, bounded payload and idempotency key. Durable inbox insertion is not task acceptance. The recipient inspects passively, checks its current authorization under client guidance, and explicitly claims the message under its runtime session. Its external host performs the work; SAGE has no task-answer oracle in this reply route. A new MCP reply checks session ownership, records completion and exposes a result for sender retrieval. Reading that labelled, potentially truncated result is another handoff, not proof of sender acceptance or delivery to the user. WF-RTE-local-exchange.
Exact retries reconcile sends and replies; conflicting content fails. Explicit handoff compares old session and revision, but does not prove the old worker is dead or undo external side effects. Session fencing is scoped to new supplied-session completion through MCP: raw calls without a session and completed replay branches differ. Optional memory judging does not grade Messages results. These mechanisms support durable coordination among existing agents, not automatic delegation acceptance, successful task execution or universal ownership enforcement. WF-RTE-local-exchange.
Scenario-relative assessment and what would change it
For authorized agents sharing project experience, the source supports governed retention, several later-context routes, conditional evidence-relative admission and recoverable local messaging. For a scenario requiring universally verified knowledge, full conversation restoration or measured self-improvement, the same source does not establish the required guarantees. Concrete counterpaths—not just missing runs—prevent the universal semantic-check and full-transcript claims. RT-RTE-co-commit, RT-RTE-targeted-recall, MEM-OBJ-capture-unit.
Provenance-linked candidate/evidence/verdict/recall/action traces under captured configuration would establish activation more strongly. Full bridge input/output pairs could resolve transformation and criticism questions; later capacity evidence and an appropriate intervention could support benefit attribution at the treatment actually varied. Benchmark source already wires retained context into answer generation and gold-relative grading, but no inspected results establish effectiveness, and this different benchmark cannot verify the reported longitudinal study. Universal semantic assurance would additionally require aligned admission across bypasses, preserved delivery qualifications and evidence of evaluator reliability. Complete restoration would require both capture coverage and a host visibility contract. MEM-RTE-evaluation-readback, EPI-RTE-benchmark-check, EPI-CLM-feedback-learning.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| Re-embedding trigger inventory is incomplete; provider-ready automatic repair is acknowledged alongside the inspected manual lane. | SRC-1; MEM-RTE-reembed-index; report verification limits | Operator migration worker and verification's provider-readiness tracing | No all-re-embedding human-initiation requirement or complete automatic-maintenance-trigger comparison. Vector regeneration and later search remain supported. | Complete recorded trigger, authority, selection and continuation traces for automatic repair and other included lanes. |
| Wider consumer, recovery, removal and transport inventories remain incomplete. | SRC-1; MEM-OBJ-memory-relationships, MEM-RTE-sync-copy, MEM-RTE-lifecycle-change, RT-RTE-machinery-change, WF-RTE-local-exchange; report verification limits | Selected graph/task/journal, snapshot/recovery, federation-removal and exact-local messaging branches; update tracing distinguishes Linux operator recovery from Darwin failed-activation restoration and detached pre-startup helper recovery | No exhaustive consumer census, universal durability/recovery, global erasure or all-client ownership guarantee. Repaired Darwin paths do not establish unrestricted post-startup downgrade or complete platform coverage. | Records tracing remaining consumers, versioned recovery/removal and raw/legacy transport alternatives; appropriate fault and recovery evidence for operational guarantees. |
| Useful learning and exercised self-improvement lack independently inspected operation and attribution evidence. | SRC-1; RT-RTE-bridge-distillation, MEM-RTE-verdict-feedback, MEM-RTE-evaluation-readback, EPI-CLM-feedback-learning, RT-BAP-retained-context; report verification limits | Source-level proposal, numerical-feedback and later-input chains; attributed study disclosure; no linked execution or inspected study results | No independently established improvement, reliable model compliance, exercised self-improvement or component-specific causal effect. The study remains attributed reporting, not inspected run evidence. | Captured deployment and candidate lineage, subsequent behavior depending on updates, validated outcomes and comparison designs; disclosed study records for independent reconstruction. |
| Partial comparison inventories group distinct parts, consumers and admission mechanisms coarsely. | SRC-1; MEM-OBJ-memory-record, MEM-OBJ-memory-indexes, MEM-OBJ-review-decision, MEM-RTE-verdict-feedback, WF-RTE-local-exchange; profile verification limits | Partial prose/symbolic form, lineage, retrieval/quorum ranking, service-fence and automatic-admission groups | Valid existence findings do not support per-object, per-consumer, per-admission inventory comparisons, unit-count statistics or an exhaustive mechanism census. | Finer units with separately supported scope, consumer, control and coverage for each included mechanism. |
| Context capture, evidence lineage and host interpretation bound later reliance. | SRC-1; MEM-OBJ-capture-unit, RT-RTE-hook-context, MEM-OBJ-gate-material, RT-RTE-targeted-recall | Selected conversational text, bounded hook output, temporary support evidence and client-specific recall serialization | No full-transcript or host-window completeness, permanent replayable support dossier, universal warning preservation or reliable downstream use. | Host injection/capacity traces, capture/reconstruction coverage, retained evidence lineage and client-specific delivered-content checks; system changes where existing exclusions or deletions preclude stronger guarantees. |
No unresolved cross-report conflict was retained in the supplied reconciliation. Both supplied verifications declare no blockers; their limits are carried above. This synthesis used accepted member records only and made no frozen-source reads.