GBrain synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From proposed content to retained memory
GBrain supplies explicit memory to existing agents through a CLI, library, MCP tools and host integrations. It also ships internal model-driven jobs. Its declared work is retaining facts with sources, supporting correction and withdrawal, and making retained material available across agents. This analysis covers the shipped repository at commit da434385572e0dc0d825649179aabed4107d6e60, with an applicability cutoff of 2026-10-11. The boundary is complete artifact, partial loop: representative producers, admission controls, consumers, maintenance and evaluators were inspected, not every shipped path. Evidence is code-grounded, with prompts and public claims treated separately as doctrine. There are no observed runs or causal experiments. External hosts, private corpora and provider internals are not independently assessed. See the boundary and evidence register and RT-CLM-memory-control.
A capture request does not become a completed write merely because it was accepted. The caller proposes content; software prepares and admits durable intent; a persistence consumer publishes it and records a receipt. Request UUIDs bind replay to the same intent. Admission and publication separately check source incarnation and authority; publication also checks page identity, revision and canonical bytes. Pending responses retain an inspection/retry identity without inventing success. File publication has a recovery protocol rather than one atomic file/database transaction. A committed receipt does not certify completion of separately queued derived effects. These are implemented controls on the inspected journaled paths, not universal guarantees covering raw filesystem or engine mutation. SRC-1 src/core/persistence/coordinator.ts, src/core/persistence/journal.ts; RT-RTE-capture-write, RT-OBJ-persistence-requests.
Remembering a fact adds a different decision: insert, duplicate, supersede, contest, hold or refuse. It requires provenance, but neither that string nor assigned confidence 1 certifies truth. Explicit remember defaults to world visibility; extraction instead resolves an unspecified visibility fail-closed to private. Facts live in SQL and may also appear in an entity-page facts fence. Subjectless or unresolved facts and other fallback branches can be database-only; Markdown is not every fact's canonical representation. Claim text, eligibility metadata, dense embeddings and fence fields have different consumers. Fence reconstruction can lose session IDs, so session-filtered recall is not established across rebuilds. SRC-1 src/core/persistence/memory-prepare.ts, src/core/facts/backstop.ts, src/core/ops/facts.ts; MEM-RTE-remember-write, MEM-RTE-extraction-write, MEM-OBJ-fact-metadata, MEM-OBJ-facts-fence.
From retained state to a later consumer
Requested query and recall supply selected evidence, not a demonstrated correct final answer. Recall combines fact-selection arms and optional page search, with eligibility, redaction and budget packing. Source grants and remote/private restrictions constrain delivery. Detailed ranking and exact enclosing-model token accounting remain incomplete. SRC-1 src/core/ops/search.ts, src/core/ops/facts.ts; RT-RTE-query-read, MEM-RTE-recall-delivery.
GBrain also supplies context without waiting for a model-originated retrieval request. SessionStart hooks can inject a context pack, and the OpenClaw assembly callback returns original messages plus a system-prompt addition. Designated core pages are a separate selector: core is disabled by default; when enabled, source eligibility, withdrawal, sensitivity, priority, page caps and a character budget determine delivery. Owner-only designation does not imply prior human approval of every edit, since notify policy permits publication followed by notice. Packs reserve core before packing other items, so a small requested token budget is not a universal cap on the complete result. These paths supply knowledge and directions through model interpretation, not deterministic model obedience. GBrain delegates actual OpenClaw compaction to the host. SRC-1 src/core/core-memory.ts, src/core/context-engine.ts, src/core/ops/facts.ts; MEM-RTE-core-delivery, MEM-RTE-context-pack, RT-RTE-context-assembly, RT-BAP-context-addition.
Compaction banking connects a write side to this later delivery. GBrain renders and redacts a bounded host-turn window into a file corpus before extraction. Consent, capability, source and kill-switch gates can leave only banked text. Automatic extraction can produce durable facts and fences. Software verifies entity-page existence before adding capped SQL manifest links; later assembly renders at most ten pointers and instructions to re-pull pages. It does not deliver the raw corpus or guarantee full-page retrieval. FIFO harvesting retains a receipt that can repair failed manifest publication without re-extraction. Inline fallback can mark the corpus ingested even after link publication fails, so retained facts need not yield a future pointer. Corpus cleanup and SQL manifest cleanup have separate horizons and unverified scheduling. SRC-1 src/core/context/checkpoint-harvest.ts, src/core/context/session-state.ts, src/core/context/corpus-segments.ts; MEM-RTE-checkpoint-harvest, MEM-OBJ-checkpoint-corpus, MEM-OBJ-checkpoint-manifest.
Withdrawal changes future eligibility rather than promising complete erasure. Forget expires matching facts, retains source/visibility/subject-scoped fingerprints, and invalidates projections. Writers can refuse matching re-admission; canonical overlays and core reads suppress stale fence content. Differently worded claims, semantic review, backups and previously delivered bytes prevent a universal deletion conclusion. Owner confirmation is another operation: it raises trust tier under permission or interactive confirmation, without enforcing an evidence-consuming factual acceptance criterion. SRC-1 src/core/facts/withdrawal.ts, src/core/facts/withdrawal-overlay.ts, src/core/trust/confirm.ts; MEM-RTE-fact-withdrawal, EPI-RTE-memory-withdrawal, EPI-RTE-owner-confirmation.
Checking generated content and changing later selection
GBrain has consequential source-relative checks, not merely provenance labels. Dream synthesis requests transcript-grounded pages. After child writes, mechanical checks can repair quoted spans, remove failed units and retain rejected text separately for review. Their scope is bounded lexical, numerical and attribution support; unchanged normalized units and exemptions narrow coverage. Optional model grounding asks whether selected excerpts state or directly imply a claim. Removal requires calibrated acting policy, adequate coverage and sufficiently low support probability outside the margin. Shadow mode, weak coverage, drift or errors can keep a candidate without accepting it. Repair persistence rebuilds projections, but can fail open after the initial writes. This is conditional post-write correction, not unconditional certification before use or independent verification of source truth. SRC-1 src/core/cycle/synthesize-verify.ts, src/core/cycle/grounding-decide.ts; EPI-RTE-dream-candidates, EPI-RTE-mechanical-check, EPI-RTE-mechanical-disposition, EPI-RTE-grounding-check, EPI-RTE-grounding-disposition, EPI-RTE-repair-retention.
Live answers have a weaker delivery consequence: quotation matching includes the question as evidence, unmatched quoted words can still be delivered with an unverified marker, and citations outside the gathered material can produce warnings rather than rejection. Matching establishes reproduction from supplied text, not endorsement. Maintenance likewise has heterogeneous transformations: the inspected fact-consolidation phase selects an existing highest-confidence claim and aggregates metadata; its name does not establish novel synthesis or truth. SRC-1 src/core/think/index.ts, src/core/cycle/synthesize-verify.ts, src/core/cycle/phases/consolidate.ts; EPI-RTE-answer-check, EPI-RTE-answer-delivery, MEM-RTE-facts-consolidation.
The strongest supported learning-related contribution is a durable feedback-to-ranking path, alongside trace-fed context production. A 1–5 retrieval rating is admitted only under feature, ownership, teaching-authority, revision, rate-limit and idempotence checks. Software derives localized weights; later hybrid ranking reads page weights and changes ordering keys after reranking without replacing raw scores. Disabled or zero influence and read failure bypass the effect. Changed content hashes damp deviation toward neutral. Events and rating history can be pruned while weights survive without complete attribution. This is wired behavioral adaptation, not merely logging, but no run demonstrates improved retrieval. Checkpoint harvest and transcript-fed dream retention also establish automatic trace-fed durable context with later consumers; they do not establish better capacity or novel knowledge. SRC-1 src/core/feedback/store.ts, src/core/search/feedback-boost.ts, src/core/search/hybrid/rank.ts; MEM-RTE-feedback-update, MEM-OBJ-ranking-weights, MEM-BAP-ranking-weights, MEM-RTE-checkpoint-harvest, EPI-RTE-repair-retention.
These findings support distinct conclusions, not a ladder. Improved capacity for future action is unestablished. No accepted record demonstrates that content-directed criticism of a consumed theory improved that capacity. The runtime does not establish theory-builder conditions 1–4: localized theory content, decisions depending on it, formulated content-directed criticism, and retained criticism shaping another round each remain unestablished as theory-builder findings. Numeric ratings and copying a representative claim do not fill those gaps. Reflection is independently unestablished: recorded operational state alone does not establish the required two-way self-representation. Computation performs the inspected internal feedback update and delegated execution, but autonomous theory building and whole-system autonomy remain unestablished. These are coverage limits, not evidenced system-wide absences. See RT-RTE-maintenance-run, RT-RTE-subagent-loop, MEM-RTE-feedback-update and the epistemic conclusion.
For self-improvement, the feedback path supplies a bounded dispositional contribution: admitted retrieval experience determines retained weights in GBrain's own later selector, directed toward favoring better-rated evidence. Its consumer, SQL channel and numerical ranking force are implemented. Over the inspected implementation horizon, no subsequent operation exercises that change in supplied evidence; exercised self-improvement and successful improvement therefore remain unestablished. The update also retains no content-directed criticism explaining why evidence failed. Parameter training is not established by this path or by configurable external chat, embedding and reranking interfaces. SRC-1 src/core/feedback/rate.ts, src/core/search/feedback-boost.ts, src/core/ai/gateway.ts; MEM-RTE-feedback-update, MEM-BAP-ranking-weights, RT-CMP-chat-model, RT-CMP-embedding-model, RT-CMP-reranker-model.
A connected delegated task, and what would change the assessment
The representative multi-step workflow is an authenticated client's gateway-enabled subagent job. Submission queues the prompt, owner identity and delegation snapshot. The worker intersects that snapshot with live grants; the model receives selected tools, guidance and recovered history. Model tool proposals pass through durable assistant/pending-execution callbacks before registered handlers act. Journaled writes wait for settlement. Resume reuses settled outcomes and refuses uncertain non-idempotent repetition rather than promising exactly-once execution. The worker owns retry/dead disposition; token-fenced completion stores the result, and owner-filtered polling delivers it. Gateway selection is conditional, generic gateway callbacks are optional, and the selected caller supplies them. A completed job can contain a refused or limited stop reason. Supplemental write accounting is not semantic review, and this submission does not require writes. SRC-1 src/core/ops/jobs.ts, src/core/minions/handlers/subagent.ts, src/core/minions/worker.ts; WF-RTE-agent-submission, RT-RTE-subagent-loop, WF-RTE-agent-delivery.
For explicit memory reuse, the account supports controlled publication, selected delivery and bounded withdrawal. For continuity after compaction, it supports banking and later pointers, not lossless reconstruction. For delegated action, it supports scoped execution and replay controls, not verified task success. For evidence-grounded answers, it supports conditional source-relative repair, not blanket factual warrant. Shared-skill revision similarly changes canonical files and catalog heads under policy/CAS checks, but does not establish installed host activation or better behavior. SRC-1 src/core/shared-skills/publication.ts; RT-RTE-skill-publication. The report verification accepts this bounded implementation account; the profile verification preserves partial coverage rather than complete mechanism comparisons.
Assessment would change with linked traces showing admitted writes reaching specific later consumers, checkpoint failures recovering under each branch, and delegated retries reaching owner-visible results. A controlled comparison varying feedback influence could establish exercised ranking dependence and, with an independent quality outcome, benefit. Source-linked candidate outputs and evidence-consuming acceptance decisions would resolve whether dream production creates and accepts non-entailed claims. BrainBench results must distinguish injected-gold plumbing tests from real-extractor mode; neither an unlabelled score nor maintenance health counters would establish semantic improvement. SRC-1 src/eval/brainbench/metrics/write-back.ts; EPI-RTE-writeback-evaluation. Complete producer/consumer records for optimization, compiled context and other unresolved paths could change the broader learning, theory and autonomy assessment. No source reads were performed for this synthesis; supporting passages remain in the cited accepted records.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| The profile omits retrieval-event/history pruning as a separately classified curation transformation. Its existing decay finding remains supported, and its inventory remains partial. | SRC-1; MEM-OBJ-retrieval-events, MEM-OBJ-retrieval-feedback, MEM-OBJ-ranking-weights; profile-verification limits | Recorded age-based event deletion, cascading feedback-history deletion and independently surviving weights | Complete mechanism-level cleanup inventory and forgetting comparisons across attribution, feedback history and weights; cleanup scheduling is also not established | A revised comparison explicitly classifying the recorded deletion mechanism, plus caller/configuration evidence for scheduling and deployed horizons |
| No observed run or causal experiment is available. | SRC-1; RT-RTE-capture-write, MEM-RTE-feedback-update, WF-RTE-agent-delivery, EPI-RTE-writeback-evaluation | Frozen implementation and doctrine, not target execution | Exercised recovery, host uptake, successful tasks, better retrieval and exercised or successful self-improvement | Provenance-linked execution traces and objective-relative comparisons isolating the changed pathway |
| Whole-loop coverage, initialization provenance and several later consumers remain incomplete. | SRC-1; RT-RTE-maintenance-run, MEM-RTE-context-pack, RT-RTE-skill-publication, EPI-RTE-dream-candidates, report verification | Selected shipped paths; unresolved pinned answers, compiled context, connectors/drains, patterns, concepts, optimization, other evaluators and provider state | Complete authority/update inventory, universal admission guarantees, system-wide theory-builder or autonomy verdict, and attribution of host responsibilities to GBrain | Accepted records tracing each material producer, admission, retained part and consumer, with separately scoped theory and self-representation evidence |
| Checkpoint banking, fact retention and pointer survival have distinct failure and cleanup contracts. | SRC-1; MEM-RTE-checkpoint-harvest, MEM-OBJ-checkpoint-corpus, MEM-OBJ-checkpoint-manifest, MEM-OBJ-fact-metadata | Bounded rendered/redacted host window, FIFO receipt repair, inline fallback, capped manifest and fence reconstruction | Lossless history, guaranteed surviving retrievable pointers, complete cleanup, cross-rebuild session recall or fully classified trace origins | End-to-end origin/rendering and rebuild records; failure/recovery traces for both harvest branches; GC scheduling and consumer refresh evidence |
| Grounding and answer checks have limited domains and conditional force. | SRC-1; EPI-RTE-grounding-check, EPI-RTE-grounding-disposition, EPI-RTE-repair-retention, EPI-RTE-answer-delivery | Selected source windows, mechanical classes, post-write repair, shadow/error paths and advisory live-answer warnings | Unconditional pre-use certification, source truth, complete proposition support or acceptance/integration of novel knowledge | Complete check/disposition consumer records, linked candidate/evidence decisions, validated judge performance and evidence of effective gating before consequential use |
| Withdrawal is scoped suppression, not universal semantic or physical erasure. | SRC-1; MEM-RTE-fact-withdrawal, EPI-RTE-memory-withdrawal | Fingerprint ledger, expiry, inspected projection invalidation and canonical overlays; purge and paraphrase review incomplete | All-consumer suppression, paraphrase equivalence, purge residual guarantees or recall of already delivered bytes | Full semantic-review, purge-verifier and propagation records covering every retained projection and relevant consumer |
| Ranking adaptation has incomplete attribution horizons and alternative consumers. | SRC-1; MEM-OBJ-retrieval-events, MEM-OBJ-retrieval-feedback, MEM-OBJ-ranking-weights, MEM-BAP-ranking-weights | Explicit admitted ratings and later page-weight hybrid stage; link consumers, implicit signals, reset authority and cleanup scheduling untraced | Complete learning-route enumeration, durable reasons for updates, complete attribution after pruning or equivalent force on link weights | Records for implicit/link pathways and reset/cleanup callers, retained update reasons with later readers, and linked ranking execution evidence |