Gnosis synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
Gnosis implements cross-session repository-context retention and consultation, not an enclosing coding-agent runtime. It preserves decisions, rejected alternatives, constraints and reasons that a later agent might otherwise lose. This account covers the whole shipped knowledge system at commit 429528bf98e424b40837d2dbe3631d0fb42da6fd, with an analysis cutoff of 2026-10-15: the gn CLI, storage, search, maintenance, embedded instructions and shipped caller guidance. Its evidence is inspected implementation, doctrine/design and attributed retained examples from SRC-1, not an observed task, passing test run or causal experiment. The external agent's model, shell authorization, context management and actual decisions remain outside the boundary. See the boundary and evidence and Source register.
Consult earlier reasons before implementation
The representative workflow is explicitly joined by the shipped planning and review instructions. Setup asks the external agent to request gn help plan at task start. Detailed planning doctrine places search after an initial plan and before implementation. The agent chooses keywords, requests search, interprets returned decisions and is instructed to surface conflicts to the user before proceeding. It later records decisions during work and requests review before committing. Gnosis prints guidance and command results; the same external caller owns each transition, conflict judgment, retry and final coding-task delivery. There is no internally managed delegation or product acceptance gate in this workflow. See the workflow account, RT-RTE-doctrine-delivery and RT-BAP-agent-guidance.
A requested search resolves the local repository store, refreshes a disposable SQLite FTS5 index when indicated by file modification time, and returns lexical matches ordered by rank. The default maximum is 20 hits. Search indexes text and topics, not author, related IDs or timestamps; its 12-token snippets can come from topics rather than explanatory prose. latest instead returns recent-by-creation entries with 200-rune snippets. show can reload the full body and metadata by ID prefix or topic. Full-entry expansion is available, not automatic or mandatory. Thus Gnosis controls requested output selection, not eviction or compaction of the host's accumulated model context. See RT-RTE-knowledge-read, MEM-OBJ-entries-fts and EPI-RTE-lexical-selection.
This distinction matters when an agent revisits a rejected design: retained reasons may avoid redundant investigation or enable a new objection, as the product claims, but only if the query finds the entry and the caller reads enough of it. A search miss is not evidence that no relevant decision exists. Lexical rank is not credibility, and a snippet can omit the condition that qualifies its visible text. See EPI-CLM-knowledge-continuity and MEM-OBJ-entry-text.
Record selected knowledge, not a verified session history
During work, doctrine requests immediate recording of decisions and explicit rejections before session context is lost. Review narrows what belongs: human context or otherwise unreconstructable empirical observations, enough explanation for later readers, and no obvious code-comment home. Conclusions reproducible from code or documentation should be omitted. These are caller-applied usefulness criteria, not executable truth or origin checks. See RT-OBJ-doctrine, EPI-RTE-candidate-filter and EPI-RTE-candidate-choice.
The caller submits selected prose. gn write checks argument shape, nonempty text, normalized topics and related-reference resolution, then appends an entry to .gnosis/entries.jsonl and returns its ID. The durable body is natural language; IDs, topics, times and references support fixed lookup and organization rules. Attribution is a caller string or Git-config fallback, not authentication. The shipped ten-entry corpus shows rationale-bearing candidates are available, but does not identify the original observation, human approval or acceptance history of each claim. See RT-RTE-entry-admission, MEM-OBJ-entry-text, MEM-OBJ-entry-metadata and EPI-RTE-entry-import.
Structural checks can veto malformed submissions; they do not license reliance on the body's truth. The inspected system has no target-implemented evidence-consuming truth-acceptance route followed by integration of an accepted claim. This bounded absence does not exclude judgments made privately by external agents or humans. Topic counts are a separate, narrow derivation over decoded topic occurrences, not verification of repository claims. See EPI-RTE-structure-check, EPI-ABS-evidence-acceptance and EPI-RTE-topic-count.
Maintain the corpus and distinguish withdrawal surfaces
Review asks the caller to remove stale entries or write a related replacement explaining the change. Edit revises retained content; removal deletes selected corpus members. A related link alone neither suppresses the old entry nor establishes precedence. Prior bodies and withdrawal reasons have no built-in history; restoration depends on another edit or external version control. Sharing likewise depends on the caller committing and distributing .gnosis, not a Gnosis synchronization service. See RT-RTE-entry-maintenance, EPI-RTE-entry-replacement and EPI-RTE-entry-withdrawal.
The storage protocol has real but limited guarantees. Cooperative production writes lock ID generation and append; maintenance rereads under an exclusive lock, preserving unrelated additions. It does not prevent an old editor buffer from overwriting a newer edit to the same entry. Write-reference validation precedes the append lock. Direct file edits and Git merges need not follow this protocol. Failure is also stage-dependent: a post-append reread can fail after retention succeeded, so retrying an unsuccessful invocation is not always retrying an unapplied write. See RT-RTE-entry-admission and RT-RTE-entry-maintenance.
Ordinary successful refresh propagates corpus changes into search, but freshness is not semantic currency or universal revocation. Equal modification times can hide changed contents; disappearance of JSONL can leave old cached hits searchable. Reindex forces rebuilding only if database opening and source decoding succeed. Neither removal nor rebuilding retracts text already delivered into an external agent's context. See MEM-OBJ-metadata and EPI-RTE-cache-refresh.
What the retained progression establishes
The strongest supported contribution is a working implementation path for preserving inspectable rationale and making it requestable in later sessions, with explicit revision and deletion capabilities. Whether criticism of a consumed account improved capacity for future action remains unestablished: no trace connects an entry, a later decision, criticism, a retained update and subsequent changed behavior. Caller-mediated session extraction is afforded, but automatic per-content admission and original trace inputs are unresolved. Automatic indexing and generated metadata do not resolve that question. This is neither a positive automatic trace-learning finding nor a system-wide negative. See MEM-OBJ-entry-text, RT-RTE-entry-admission, RT-BAP-agent-guidance and the independent profile verification.
The accepted reports do not establish theory-builder conditions 1–4 as a completed loop. Rationale-bearing text is inspectable, but the records do not supply separately supported condition assessments that demonstrate content-dependent consumption, content-directed criticism and retention shaping another round. Reflection and autonomous theory building are independently unestablished, not inferred from persistence or instruction delivery. Exercised self-improvement at this shipped-system boundary is also unestablished: no evidence closes an objective-bearing criticism/update pathway through changed organization to causally dependent later operation. These are limits on attribution, not a grade of the knowledge CLI. See the epistemic bounded conclusion and RT-BAP-agent-guidance.
For a task requiring earlier repository intent, the implementation offers durable reasons and explicit consultation, conditional on caller selection and use. For a task requiring verified claims, guaranteed conflict prevention or assured withdrawal from all consumers, these routes provide no such guarantee. Original session and decision traces would resolve provenance and admission control; linked later-consumer traces would test reliance and criticism; outcome comparisons with independently varied retrieval or review would support narrower benefit attribution. Version-aware edits, typed supersession and stronger cache invalidation would change specific operational limits, not by themselves establish semantic truth or improved decisions. The report verification and profile verification declare no additional limits or blockers. This synthesis adds no source inspection or execution evidence.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| No observed execution or component comparison | SRC-1; RT-RTE-knowledge-read, EPI-RTE-test-check, RT-CLM-agent-compliance | CLI implementation, selected test source and shipped instructions | Actual compliance, test passage, deployment reliability, retrieval quality and causal task benefit | Inspectable command/test and agent traces; outcome comparisons varying the claimed component |
| Original provenance and caller admission control unresolved | SRC-1; MEM-OBJ-entry-text, MEM-OBJ-entry-metadata, EPI-RTE-entry-import | Shipped corpus, caller input and attribution paths | Authenticated human origin, complete lineage, manual versus automatic prose admission and qualifying trace-learning source | Original evidence linked to entries and per-content composition, approval and update records |
| Later reliance and criticism-to-update progression unobserved | SRC-1; RT-BAP-agent-guidance, EPI-CLM-knowledge-continuity, EPI-RTE-entry-replacement | Requested delivery, retained reasons and explicit maintenance | Improved future capacity, completed theory-building iteration, reflection, autonomous theory building or exercised self-improvement | Linked consumption, criticism, update and later-operation evidence, with each property assessed independently |
| Semantic judgment belongs to the excluded host | SRC-1; EPI-ABS-evidence-acceptance, RT-BAP-agent-guidance | Dispatched commands and embedded doctrine, not private agent/human reasoning | Claims that retained content is verified, or that external users never perform evidence-consuming acceptance | Evidence of the actual external evaluator, criterion, evidence, decision and consequential consumer |
| Requested output can omit reasons or qualifications | SRC-1; EPI-RTE-lexical-selection, MEM-OBJ-entries-fts | Sanitization, lexical selection, limits, snippets and full-entry alternatives | Exhaustive conflict detection or complete rationale delivery from a successful search | Query/result and full-read traces, plus relevance and qualification-preservation evaluation |
| Maintenance and withdrawal guarantees are path-specific | SRC-1; RT-RTE-entry-maintenance, MEM-OBJ-metadata, EPI-RTE-entry-withdrawal | Cooperative locks, edit branches, untyped references and mtime cache refresh | Universal lost-update prevention, enforced supersession or complete withdrawal from caches and prior host context | Revised version/invalidation controls and evidence covering alternate writes and each later consumer |
| External infrastructure and optional helper coverage remain bounded | SRC-1; boundary coverage, RT-RTE-entry-maintenance | Shipped knowledge loop; dependencies, release machinery and some helper/test bodies excluded or not exhaustively inspected | Exhaustive portability, concurrency, rendering, release-reliability or test-adequacy guarantees | Relevant helper/dependency and release inspection, with appropriate execution evidence |