Hippocampus synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
A host-operated knowledge vault
Hippocampus supplies a personal Markdown knowledge system, not an enclosing agent executor. An external Claude Code or Codex agent interprets its project contract and ingest, save and maintenance skills. Host-native tools perform file and web operations; local shell/Python utilities provide lifecycle bookkeeping, structural checking and framework copying. Obsidian is an optional browsing consumer. The claimed work is to turn messy sources and conversation conclusions into reusable, linked pages and retrieve them in later sessions without a separate vector retrieval service (RT-OBJ-project-contract, RT-RTE-source-ingest, RT-RTE-conversation-save, RT-RTE-vault-query).
The evidence is SRC-1, the repository at 99531951ae60a9814ac7d27a32dd00f0c81a5008, with an applicability cutoff of 2026-11-08. The whole-system boundary covers shipped producers, maintenance, persistence, consumers and evaluators, not external host/model implementations or populated private vaults. It is code-grounded for utilities and configuration; semantic workflows remain instruction-grounded. The repository supplies an empty content scaffold, not execution traces or experiments. Thus implemented checks and configured delivery can be distinguished from requested model behavior, but successful end-to-end operation and comparative benefit are unobserved (boundary, Source register, report verification).
From a source or conversation to retained pages
The representative workflow is a user asking the foreground agent to ingest one local inbox note. The agent reads the source, computes its hash and searches source-page hashes. A duplicate requests stopping unless the user explicitly asks for re-ingest. It then reads recent context and the catalog, chooses a noncolliding archive path, creates a templated source page and selectively creates or updates reusable entity/concept pages. It adds links, flags contradictions on both sides, updates catalog/journal/cache, moves the original, and reports pages changed, contradictions and an insight. The model owns semantic progression and most content admission. No child worker, isolated extraction context or independent reviewer is specified. Batch processing is instructed to be sequential, not protected by a lock (RT-RTE-source-ingest, workflow account).
This separates three retained layers: original material, a source account with hash/origin metadata, and reusable knowledge pages. Local originals are to be preserved without editing; outside files are copied, pasted text saved verbatim, and URLs archived as fetched extraction with metadata—not necessarily raw response bytes. Very large sources permit section-wise provisional extraction and final consolidation. Those exceptions matter: complete reading is a model protocol, not a lossless acquisition guarantee. Ordinary writes need no universal per-page human approval; new ongoing projects require confirmation, and contradictory claims await user resolution (MEM-OBJ-archive, MEM-OBJ-source, MEM-OBJ-pages).
Conversation save offers a second write path. It asks the model to retain conclusions, decisions/rationale or findings rather than the transcript, with selective questions for ambiguity or same-name overwrites. This is a supported partial contribution: session experience can be transformed automatically into durable pages and derivative cache content that later agents can consume. The route is afforded by instructions, not observed in a supplied run. It establishes a trace-fed retention pathway, not faithful preservation of every reason, parameter learning or improved capacity for future action (RT-RTE-conversation-save, memory annotation, profile verification).
Delivery, persistence and later consumption
The immediate ingest delivery is a user report following the requested file/archive/metadata changes. A Git commit is a subsequent lifecycle action, not a content-acceptance criterion. At an eligible first Stop, if hot exists and another wiki Markdown file has a newer modification time, the shell emits a blocking request for the agent to rewrite hot. The host must honor the request and the model must produce the summary. Template markers disable automation; missing hot, recursive Stop and equal/older timestamps bypass the freshness guard. This is a bounded refresh reminder, not semantic validation or an invariant that every changed claim appears in recent context (RT-RTE-lifecycle-hooks, RT-CLM-cache-refresh, RT-BAP-stop-refresh).
When persistence is reached, the hook stages content roots and attempts a local commit. The commit consumes the entire existing index, including unrelated pre-staged files; restricting the hook's own staging does not restrict that commit. Failure is logged and still returns zero. Multi-file ingest has no supplied transaction or automatic rollback, so originals and successful Git history offer conditional manual recovery rather than atomic delivery (RT-RTE-lifecycle-hooks, workflow recovery account).
Later consumption has two routes. On startup/resume/clear/compact events, configured hooks print the whole existing hot file without question-targeted selection. In requested answering, the agent reads hot, then catalog descriptions, chooses three to five pages, follows at most one further link hop, and if still uncertain searches distinctive terms, aliases and synonyms. Search can recover unindexed content, so catalog omission does not imply literal invisibility. Answers are instructed to cite vault pages or report a scoped knowledge gap after fallback, rather than substitute training knowledge. The cross-project consultation snippet affords another reader role without demonstrating deployed wiring (MEM-OBJ-hot, MEM-OBJ-index, RT-RTE-vault-query, MEM-BAP-vault-knowledge).
The working-context strategy is selective external reading and summary reload, not a local implementation of conversation trimming or isolated child frames. The 500-word hot policy is checked by a warning, not enforced at delivery; Codex's configured context limit has external semantics. Pages omitted from context remain reachable by explicit reads, but links are strings, not automatically followed live references. This makes a stack-frame account of delegated work inapplicable to the specified foreground workflow (RT-RTE-lifecycle-hooks, memory read-back account).
What checking licenses—and what it does not
Maintenance combines an executable structural scanner with instructed editorial judgment. Mechanical findings concern parsed metadata, filenames, links and counts; editorial findings allege stale claims, missing links/pages or aging seeds, with sampling allowed in large vaults. They have different evaluators and warrant. The scanner's lenient parser, basename resolution and removal of heading fragments limit what its findings establish. It neither verifies world claims nor detects every semantic duplicate or missing target heading. Completed scans return zero even with error findings (EPI-OBJ-mechanical-findings, EPI-OBJ-editorial-findings, EPI-CLM-structural-checks).
The lint skill asks for bulk approval of safe fixes and individual approval of merges, deletions and contradiction resolution, then edits and rescans. That policy does not intercept ordinary ingest/save writes or Stop commits. Contradiction flags preserve competing candidates rather than retract either side or force later abstention. Source-distrust and citation rules likewise depend on model/host compliance; they are not demonstrated confinement or injection resistance (RT-RTE-vault-lint, RT-RTE-source-ingest).
Hashes, origins, confidence labels, maturity status, citations and graph links provide identity, attribution or organization. None alone records evidence-consuming acceptance against a truth criterion for a named use and scope. Source/save/summary transformations remain semantically indeterminate without paired inputs and outputs. The records support narrow program-derived structural claims, but not produced accepted ampliative knowledge or post-acceptance lifecycle integration merely from filing linked pages. The design argument that deliberate links encode meaningful intent does not establish that the intent is true or retrieval outperforms similarity search (epistemic conclusion, EPI-CLM-graph-knowledge).
Change and assessment within this boundary
Framework synchronization is a separate behavior-shaping change surface. Update copies eligible incoming framework files after a dirty-file veto for differing existing destinations; export copies eligible local files into a destination without that veto. Neither route establishes semantic quality, loaded deployment or an evidence-responsive improved successor. The export path therefore defeats a universal no-overwrite reading without demonstrating actual data loss (RT-RTE-framework-update, RT-RTE-framework-export, RT-CLM-local-edit-protection).
For the intended personal-vault scenario, the supported contribution is inspectable retained text, explicit provenance pointers, selective reads, recent-context reload and localized structural diagnosis. Catalog selection assumes an index that fits model context; the source acknowledges wording misses and scale limits. There is no measured recall, latency, scalability or superiority result. The model is external and not pinned, with provider mutability uninspected (MEM-OBJ-index, EPI-CLM-graph-knowledge, RT-CMP-agent-model).
The accepted runtime records do not provide a condition-by-condition theory-builder assessment, so this synthesis does not infer one from templates that permit explanations and rationale. Nor does the afforded conversation-to-memory chain demonstrate that criticism of a consumed theory improved future capacity. Reflection and autonomous theory building remain unestablished independently: an operations journal is not by itself evidence of reciprocal self-representation, and model-directed ordinary writes do not establish computational ownership of every theory-building role. Exercised self-improvement is also unestablished: the records show potential later knowledge delivery and framework replacement, but no objective-bearing evidence-to-update-to-later-behavior causal chain. These limits do not erase the afforded trace-fed write contribution or imply absence of hidden model reasoning (MEM-OBJ-pages, MEM-OBJ-log, RT-RTE-conversation-save, epistemic conclusion).
Evidence that would change this assessment includes provenance-linked source/conversation-to-page pairs, actual host delivery and later query traces, criterion-bound contradiction decisions with derivative corrections, and outcome comparisons at the level of the claimed mechanism. A component benefit would require independent variation of that component. Enforced semantic freshness, transactional ingest or destination-dirty checks on export would change the corresponding control assessment; mere successful delivery or copying would not establish those stronger guarantees.
Limitations
| limitation | affected SRC IDs and record citations | inspected boundary | conclusion prevented | evidence that would resolve it |
|---|---|---|---|---|
| No supplied operation or outcome evidence | SRC-1; RT-RTE-source-ingest, RT-RTE-vault-query, RT-CMP-agent-model | Shipped framework and empty vault scaffold; external host/model internals excluded | Successful model compliance, activated memory, injection resistance, recall, latency, scalability or comparative benefit | Provenance-linked runs with host/model identity, actual delivered context and outputs; scoped evaluations and suitable comparisons |
| Exact host-visible cache and compaction survivors are unknown | SRC-1; MEM-OBJ-hot, RT-RTE-lifecycle-hooks | Complete hook configuration and shell; external context-limit/compaction semantics unavailable | Exact token budget, complete delivered summary or restoration of discarded conversation | Versioned host semantics and inspectable pre/post-compaction and hook-delivery traces |
| Semantic fidelity, novelty and acceptance are unresolved | SRC-1; MEM-OBJ-source, MEM-OBJ-pages, RT-RTE-conversation-save | Templates and prescribed acquisition/save transformations; no populated paired artifacts | Lossless rationale retention, entailed or ampliative classification of a saved claim, accepted ampliative knowledge | Actual input/output lineage plus evidence-consuming acceptance identifying criterion, intended use and scope |
| Withdrawal does not specify correction of every derivative | SRC-1; MEM-OBJ-pages, MEM-OBJ-hot, RT-RTE-vault-lint | Contradiction and editorial procedures, cache replacement | Guaranteed removal of withdrawn claims from downstream answers or summaries | Explicit dependency/invalidation procedure and traces of corrected derivatives reaching later consumers |
| Arbitrary payload forms, provenance and trace origins remain unresolved | SRC-1; MEM-OBJ-archive, MEM-OBJ-log, RT-RTE-conversation-save | Empty archives/attachments and documented input branches; known conversation-save path | Complete form/lineage/trace-learning inventory or system-wide negative about other experience-fed paths | Representative actual payloads and their original provenance, transformation and later-consumer chains |
| Optional retention does not establish later consumption | SRC-1; EPI-OBJ-findings, MEM-OBJ-archive, MEM-OBJ-log | Optional report, journal and origin-pointer contracts | Routine later authority or automatic learning from those retained objects | A specified or observed later selector, delivered part, consumer and consequential use |
| Higher-order improvement properties are not established | SRC-1; MEM-OBJ-pages, RT-RTE-conversation-save, RT-RTE-framework-update | Accepted memory/write/copy records without a condition-specific theory account or exercised causal chain | Theory-builder conditions 1–4, criticism-driven capacity improvement, reflection, autonomous theory building or exercised self-improvement | Separate property-specific records and evidence, including retained criticism and next-round consumption where relevant, reciprocal self-representation for reflection, and objective-linked change exercised in later behavior for self-improvement |
The report and profile verifications declare no additional limits or blockers. No unresolved cross-report conflict remains in the supplied reconciliation. This synthesis used the supplied members only; no frozen source reread was needed (report verification, profile verification).