Binder synthesis

Type: agentic-system-analyses/types/agentic-system-synthesis.md

Bounded synthesis

From caller input to shared state

Binder is a local-first memory and knowledge system, not an enclosing autonomous agent runtime. It gives people, scripts and agent hosts typed persistent state, query interfaces and editable document projections. This account covers the shipped repository at commit e4066d750ac5c6918706f44b12daa261b261106e, including its engine, CLI, MCP/HTTP interfaces, editor integration, library callers, agent skills and examples. The applicability cutoff is 2026-10-15. SRC-1 supplies inspected implementation and doctrine/design, not an observed execution or causal experiment. No target tests, examples or model calls were run. External hosts, providers and unshipped extensions are not independently characterized. See the boundary and evidence and Source register.

The ordinary progression is caller input → graph admission → SQLite persistence → post-commit maintenance → later selection and consumption. CLI update, MCP transact, HTTP transactions and the bundled workspace library converge on graph update. Code controls this progression; the caller supplies the proposed content. Update reloads schema inside a reserved write transaction, processes config changes before records against the resulting schema, and checks transaction sequence and predecessor hash before application. These checks enforce structural consistency, not the truth of stored prose. Database rollback covers failed graph-managed SQL work under SQLite/adapter semantics. Prepared replay is a distinct path: it retains sequence checks but skips fresh-input preparation. See RT-RTE-graph-update, RT-CLM-transaction-boundary and RT-RTE-journal-repair.

Persistence does not imply universal approval. Built-in HTTP/MCP writes have no mandatory per-content human confirmation; author/client labels are provenance, not authentication. Some skills separately require human approval. Workspace plugins, shell hooks, scripts and custom HTTP modules can cross the graph's effect boundary, so structural admission is not a host-wide sandbox or effect guarantee. See RT-RTE-extension-load and EPI-RTE-import-mapping.

From committed state to later context

Journal, undo and document plugins maintain logs, recovery state and projections after database success. Subscriber failure does not undo that commit. Disk/editor saves travel back through extraction and graph update; outward rendering instead selects graph content through navigation and views. Verify mode can preserve diverged files, whereas force can overwrite them. Neither direction establishes lossless semantic round trips. Snapshot paths, hashes and timestamps are synchronization metadata, not semantic summaries. See RT-CLM-derived-effects, RT-RTE-document-sync, MEM-RTE-document-projection and MEM-OBJ-snapshot-metadata.

Later processes can reopen the graph and fetch retained fields. Keys, filters, relation expansion and field selection determine what reaches a consumer. This matters because a rendered parent document may contain child summaries rather than raw logs. Search defaults to 50 graph rows; CLI argv --limit trims after search rather than necessarily enlarging retrieval. These are record/field budgets, not input-token budgets or evidence of exhaustive duplicate detection. Journal skill preloading and task rereads are requested delivery, not an established always-on model-memory injector. Commit-triggered hooks and document projection provide separate non-model push paths. See MEM-RTE-selected-read, MEM-OBJ-config-views and RT-RTE-extension-load.

From activity to journal guidance

Binder's strongest supported experience-reuse contribution is an explicit retained-content chain: activity and conversation capture → day/period summaries → later briefings, review and planning. Logs, derived descriptions and future commitments remain different objects. Summary procedures preserve source plans and leave logs separately accessible; later field selections need not deliver all those reasons. Retained summaries supply context, while goals/plans supply directions through human or model interpretation. See MEM-OBJ-journal-log, MEM-OBJ-journal-summary, MEM-OBJ-journal-plan and MEM-BAP-journal-guidance.

The semantic procedure is shipped prompt policy, not an automatic database summarizer. Day/period review asks for evidence-grounded proposals, user corrections and confirmation for personal journal/planning use. Other capture paths instruct agent-written compact conversation records without separate per-content approval. Those automatic trace-fed capture paths support trace learning at the claimed layer: durable contextual artifacts have named later consumers. This classification establishes neither improved future capacity nor reliable extraction. Human-approved summary and planning writes retain their separate admission rules. See MEM-RTE-journal-curation, EPI-RTE-journal-review and the profile verification.

The learnings field does not itself establish new warranted knowledge. Proposed content can summarize activity, import outside information or infer patterns; without concrete outputs, its semantic relation remains indeterminate. User confirmation licenses a personal account, not a general causal explanation or transfer claim. The Git importer is narrower still: it reads commits and retained logs, asks an external agent for entries, parses them and prints a proposal without applying it. Its prompt drops commit hashes and time-of-day; timestamp positioning cannot recover that missing evidence. A permitted empty list is nevertheless rejected by its parser. See EPI-CLM-period-learnings, EPI-RTE-journal-review, RT-RTE-git-import and EPI-RTE-import-format.

From task request to intended delivery

The representative workflow is implement-task, because its entry procedure explicitly joins conditional investigation, discussion, design, testing and implementation through stored task rereads. A fix lacking a root cause first requests reproduction and investigation. The same foreground agent stores diagnosis, reproduction steps and source paths, removes temporary investigation edits, rereads the task and selects the next guide. This is a retained handoff, not delegation to an isolated worker. Diagnosis can influence later work without a separate content-approval step, supporting a claimed trace-fed retention pathway; reproducing the symptom does not uniquely validate its proposed cause. See RT-RTE-task-guide, MEM-OBJ-task-findings and EPI-RTE-task-findings.

Discussion and design introduce explicit user agreement for Decisions and the specification. Tests have phase-dependent expectations: a preimplementation reproduction test should fail, whereas final verification should pass. Implementation prioritizes existing passing tests, specification, acceptance criteria and then potential solution; failures are assigned back to the foreground agent to fix. These are prompt-level completion conditions, not an executable release gate. Intended delivery is in-scope code meeting criteria and passing tests, with unrelated changes excluded. Task completion status, commit, merge, release, final human acceptance and mechanical termination are not established by that checklist. See MEM-OBJ-task-specification, RT-BAP-task-guidance and the workflow account and annotations.

What the supported paths establish

For scenarios needing typed state across invocations, Binder has wired persistence, structural admission and selected read-back. For journal review or task work, it additionally supplies concrete host-interpreted procedures for evidence use, correction, retained guidance and repair. The evidence does not establish reliable semantic curation, successful delivery or improved future capacity. No comparison measured an improvement, much less isolated its cause. The positive trace-fed retention paths stand independently of that unresolved benefit question.

The reports do not establish a complete theory-builder assessment of conditions 1–4: retained explanatory fields and requested consumption are relevant contributions, but cannot substitute for separately evidenced content-directed criticism and iteration. In particular the discussion-object inventory remains incomplete. Criticism of a consumed theory improving future capacity is unestablished. Reflection-themed journal work does not alone establish reflection of Binder's own organization; autonomous theory building is likewise unestablished, independently of deterministic engine automation. At this boundary, exercised self-improvement is not demonstrated: task code can be an external work product, while retained plans can shape later operation, but no trace closes evidence-responsive self-change through subsequent behavior causally dependent on that change. These are separate unresolved properties, not a bundled negative. See EPI-RTE-journal-review, RT-RTE-task-guide and report-verification limits.

Evidence that would change these assessments includes linked capture/write/later-consumption traces; actual proposals, corrections and retained reasons; and objective-specific before/after outcomes with comparisons adequate for attribution. A complete theory-cycle comparison also needs the retained discussion objects and their consumers. Stronger synchronization or recoverability conclusions require the separate broader-doctrine assessment identified below. No source reads were added during synthesis.

Limitations

limitation	affected SRC IDs and record citations	inspected boundary	conclusion prevented	evidence that would resolve it
Discussion objects are not separately traced in the epistemic inventory; carried report-verification limit.	SRC-1; EPI-OBJ-task, EPI-RTE-task-findings, RT-RTE-task-guide; verification limit	The workflow inspected the current discussion guide and records its requested arguments and user-agreed Decisions; the epistemic inventory does not trace those retained parts and later consumers separately.	A complete comparison of argument retention, criticism, decision authority and theory-cycle organization across task guides; the supported workflow remains valid.	Accepted records tracing Problem/Option/Argument/Decision content, admission and later consumers; concrete criticism and iteration evidence for exercised-cycle claims.
Broader history and synchronization doctrine is outside the completed claim comparison; carried report-verification limit.	SRC-1; EPI-CLM-agent-memory, RT-CLM-transaction-boundary, MEM-RTE-history-maintenance; verification limit	Repository-concept doctrine additionally promises immutable append-only history, complete auditability/reversal and conflict-free multi-device sync. Reports trace local SQL consistency and explicit history replacement, not those full promises.	Complete historical recoverability or conflict-free distributed synchronization; local persistence/admission/recovery findings remain supported.	Claim-specific records tracing merge, rehash, replacement, reversal and distributed conflict paths, with appropriately scoped execution evidence.
No observed operation or causal comparison.	SRC-1; Source register, RT-RTE-graph-update, MEM-RTE-journal-curation, RT-RTE-task-guide	Source implementation, prompts and authored tests only; no live journal/task history, provider run or passing test result.	Actual compliance, accepted ampliative knowledge, exercised improvement and measured semantic/retrieval/delivery benefit.	Provenance-linked executions through writes and later consumers; outcome evidence and controlled comparisons for causal claims.
Input provenance and consumer inventory remain partial.	SRC-1; MEM-RTE-journal-curation, MEM-RTE-task-curation, MEM-RTE-selected-read, profile verification	Known journal/task transformations and read surfaces; opaque calendar/health/activity injection, live initialization and incompletely recorded app/browser/guide alternatives.	Complete memory-object, admission, lineage, authority, direction, selector or trace-source inventories; opaque inputs cannot inherit a known branch's classification.	Producer-to-payload-to-consumer records for injection and remaining alternatives, including original inputs, selection and content decisions.
Maintenance and document algorithms were not exhaustively audited.	SRC-1; MEM-RTE-history-maintenance, MEM-RTE-document-projection, RT-RTE-document-sync	Inspected command gates, snapshot orchestration, replay and bounded extraction/cleanup paths; full slot grammar, squash/rehash and backup restoration remain incomplete.	Lossless semantic round trips, preservation of every rationale/provenance field, whole-batch atomicity or universal reversal.	Algorithm-specific records and representative preservation, divergence and restoration evidence.
Host/provider and arbitrary extension behavior are outside captured operation.	SRC-1; RT-RTE-extension-load, RT-RTE-llm-call, RT-CMP-journal-model, RT-CMP-llm-model	Shipped loaders, subprocess caller and fresh-context completion wrapper, not provider internals, enabled grants or unshipped code.	Host-wide isolation/authentication, pinned model weights, provider forgetting, model-internal learning or Binder-owned compaction/frame guarantees.	Explicit host/provider contracts and scoped traces/configuration; separately bounded extension-effect evidence.